CN111191648A

CN111191648A - Method and device for image recognition based on deep learning network

Info

Publication number: CN111191648A
Application number: CN201911389478.7A
Authority: CN
Inventors: 陆舟; 于华章
Original assignee: Feitian Technologies Co Ltd
Current assignee: Feitian Technologies Co Ltd
Priority date: 2019-12-30
Filing date: 2019-12-30
Publication date: 2020-05-22
Anticipated expiration: 2039-12-30
Also published as: CN111191648B

Abstract

The invention discloses a method and a device for image recognition based on a deep learning network, and relates to the field of image recognition. The device transmits the first parameter data to a deep learning network, and identifies all pictures acquired from a picture sample library to obtain a second list; and sequentially carrying out fifth operation on the predicted frame data in the second list to obtain a fifth list: obtaining total deviation data according to all deviation data in the fifth list, comparing the total deviation data with preset deviation data, updating the first parameter data by using a fourth preset algorithm according to the total deviation data and the preset deviation data when the total deviation data is greater than or equal to the preset deviation data, and emptying the second list and the fifth list; and when the total deviation data is smaller than the preset deviation data, processing the picture to be identified by using a deep learning network containing first parameter data to obtain a prediction frame data table, and ending.

Description

Method and device for image recognition based on deep learning network

Technical Field

The invention relates to the field of image recognition, in particular to a method and a device for image recognition based on a deep learning network.

Background

Deep Learning (DL, one of the key research fields of machine Learning technology in artificial intelligence) is to implement artificial intelligence in a computing system by establishing artificial Neural Networks (ans) with a hierarchical structure. The hierarchical artificial neural network used therein has various forms, and the complexity of its hierarchy is commonly referred to as "depth"; deep learning uses data to update parameters in its construction to achieve a training goal, a process commonly referred to as "learning," which can be used to learn high dimensional data for complex structures and large samples;

however, the conventional image recognition method generally includes: manually designing an algorithm aiming at an object on a picture to be recognized; then, pictures containing the object images are transmitted into an algorithm one by one, and parameters of the algorithm are manually adjusted step by step according to the deviation condition of the images on the algorithm identification pictures until the deviation of the images meets the design requirements; the designed algorithm is only suitable for objects on pictures used in algorithm design, but not suitable for recognition of images of other objects, if other objects are to be recognized, a new algorithm needs to be designed for the other objects to be recognized again, and the generalization performance of the model is poor; and the parameters in the algorithm are adjusted repeatedly manually according to the identified deviation condition of the image on each picture, and the parameter adjusting method has high manual dependence and wastes time and labor.

With the development of social economy, huge image information is generated every day; the application of a deep learning network suitable for large sample data to the image recognition field to recognize images has become the most urgent need.

Disclosure of Invention

The invention aims to overcome the defects of the prior art and provides a method and a device for image recognition based on a deep learning network.

The invention provides a method for image recognition based on a deep learning network, which comprises the following steps:

step S0: the device transmits the first parameter data into a deep learning network, and acquires a picture sample library from a picture database;

step S1: when a picture is obtained from the picture sample library, dividing the picture into first preset data grids according to first preset data and calculating grid coordinate data of each grid; identifying the acquired picture by using a deep learning network comprising first parameter data according to first preset data, second preset data and third preset data to obtain a prediction frame data table, wherein the prediction frame data table comprises second preset data prediction frame data corresponding to all grids respectively; sequentially acquiring grids and grid coordinate data corresponding to the grids from first preset data grids, and executing a first operation on the acquired grids;

the first operation is: the device determines the image corresponding to the grid and the actual coordinate data of the image corresponding to the grid according to the grid coordinate data and the actual coordinate data table of the image corresponding to the picture; sequentially acquiring prediction frame data from second preset data prediction frame data corresponding to the grids, and executing a second operation on the acquired prediction frame data: the second operation includes steps S1-01 to S1-03:

step S1-01: the device acquires all category probability data from the prediction frame data, selects the maximum category probability data from all category probability data, and determines the category data of the prediction frame data according to the maximum category probability data;

step S1-02: the device acquires an image corresponding to the grid and actual coordinate data of the image; sequentially acquiring images and actual coordinate data of the images from the images corresponding to the grids and the actual coordinate data of the images, and executing a third operation on the acquired images and the actual coordinate data of the images; sequentially acquiring the images from the same category image table corresponding to the prediction frame, and executing a fourth operation on the acquired images:

the third operation is: the device acquires the class data of the image from the actual coordinate data of the image, and correspondingly stores the prediction frame data, the image and the actual data of the image to a prediction frame corresponding same class image table when the acquired class data of the image is the same as the class data of the prediction frame data;

the fourth operation is: when the device obtains the actual coordinate data of the image from the corresponding image table of the same category, calculating the prediction frame data and the actual coordinate data to obtain a second confidence parameter; calculating the confidence corresponding to the image according to the first confidence parameter, the eleventh preset data, the twelfth preset data and the second confidence parameter in the prediction frame data, and correspondingly storing the confidence, the prediction frame data, the category data and the actual coordinate data of the image into a first list;

step S1-03: the device obtains the maximum confidence in the first list; judging whether the maximum confidence is greater than the first comparison data, if so, correspondingly storing prediction frame data corresponding to the maximum confidence and the actual coordinate data of the image, which are acquired from the first list, to a second list, and emptying the first list and the prediction frame corresponding to the image tables of the same category; otherwise, emptying the first list and the image list of the same category corresponding to the prediction frame;

step S2: when all pictures in the picture sample library are obtained, the device sequentially obtains prediction frame data from the second list, and fifth operation is carried out on the obtained prediction frame data;

the fifth operation is: the device calculates deviation data of the prediction frame data and the actual coordinate data according to the prediction frame data and the actual coordinate data, and stores the deviation data to a fifth list;

step S3: the device obtains total deviation data according to all deviation data in the fifth list, compares the total deviation data with the preset deviation data, and executes step S4 when the total deviation data is greater than or equal to the preset deviation data; when the total deviation data is smaller than the preset deviation data, processing the picture to be identified by using a deep learning network containing first parameter data to obtain a prediction frame data table, and ending;

step S4: the apparatus updates the first parameter data using a fourth preset algorithm based on the total deviation data and the preset deviation data, clears the second list and the fifth list, and returns to step S0.

The invention also provides a device for image recognition based on the deep learning network, which comprises the following steps: the device comprises a transmitting module, an obtaining module, an executing module, a deviation comparing module, a module to be identified and a parameter updating module;

the transmitting module is used for transmitting the first parameter data into the deep learning network;

the acquisition module is used for acquiring a picture sample library from a picture database;

the acquisition module is also used for acquiring a picture from the picture sample library, dividing the picture into first preset data grids according to first preset data and calculating grid coordinate data of each grid; identifying the acquired picture by using a deep learning network comprising first parameter data according to first preset data, second preset data and third preset data to obtain a prediction frame data table, wherein the prediction frame data table comprises second preset data prediction frame data corresponding to all grids respectively; sequentially acquiring grids and grid coordinate data of the grids from first preset data grids;

the execution module is configured to execute a first operation on the mesh acquired by the acquisition module:

the execution module comprises a first execution operation unit;

the first execution operation unit comprises a determination module unit and a second execution operation unit;

the determining module unit is used for determining the image corresponding to the grid and the actual coordinate data of the image corresponding to the grid according to the grid coordinate data of the grid and the actual coordinate data table of the image corresponding to the picture acquired by the acquiring module; acquiring prediction frame data from second preset data prediction frame data corresponding to the grids in sequence, and triggering a second execution operation unit;

the second execution operation unit is configured to perform a second operation on the acquired prediction box data of the determination module unit:

the second execution operation unit comprises a first determination unit, a third execution operation unit, a second acquisition unit, a fourth execution operation unit and a judgment, storage and emptying unit;

the first determining unit is configured to obtain all category probability data from the prediction frame data obtained by the determining module unit, select the largest category probability data from the all category probability data, and determine the category data of the prediction frame data according to the largest category probability data; acquiring an image corresponding to the grid and actual coordinate data of the image; acquiring images and actual coordinate data of the images from the images corresponding to the grids and the actual coordinate data of the images in sequence, and triggering the third execution operation unit;

the third executing operation unit is configured to execute a third operation on the image acquired by the first determining unit and the actual coordinate data of the image:

the third execution operation unit is configured to acquire category data of the image from the actual coordinate data of the image acquired by the first determination unit, and when the acquired category data of the image is the same as the category data of the prediction frame data, store the prediction frame data and the image and the actual data of the image in correspondence to a prediction frame-corresponding same-category image table;

the second obtaining unit is used for sequentially obtaining images from the same-class image table corresponding to the prediction frame obtained by the third executing unit;

the fourth execution operation unit is configured to calculate the prediction frame data determined by the determination module unit and the actual coordinate data of the image acquired by the second acquisition unit to obtain a second confidence parameter; calculating the confidence corresponding to the image according to the first confidence parameter, the eleventh preset data, the twelfth preset data and the second confidence parameter in the prediction frame data, and correspondingly storing the confidence, the prediction frame data, the category data and the actual coordinate data of the image into a first list;

the judgment, storage and emptying unit is used for acquiring the maximum confidence coefficient in the first list obtained by the fourth execution operation unit; judging whether the maximum confidence is greater than the first comparison data, if so, correspondingly storing prediction frame data corresponding to the maximum confidence and the actual coordinate data of the image, which are acquired from the first list, to a second list, and emptying the first list and the prediction frame corresponding to the image tables of the same category; otherwise, emptying the first list and the image list of the same category corresponding to the prediction frame;

the acquisition module is further configured to acquire prediction frame data and actual coordinate data of a corresponding image from the second list when all the pictures in the picture sample library are acquired, and trigger the fifth execution operation unit;

the fifth execution operation unit is used for calculating deviation data of the predicted frame data and the actual coordinate data according to the predicted frame data and the actual coordinate data, and storing the deviation data to a fifth list;

the deviation comparison module is used for obtaining total deviation data according to all deviation data in the fifth list obtained by the fifth execution operation unit and comparing the total deviation data with preset deviation data;

the processing to-be-identified module is used for processing the to-be-identified picture to obtain a prediction frame data table by using a deep learning network containing first parameter data when the total deviation data obtained by the deviation comparison module is smaller than the preset deviation data, and ending;

and the parameter updating module is used for updating the first parameter data by using a fourth preset algorithm according to the total deviation data and the preset deviation data and emptying the second list and the fifth list when the deviation comparing module obtains that the total deviation data is smaller than the preset deviation data.

Compared with the prior art, the invention has the following advantages: the invention provides a method and a device for image recognition based on a deep learning network; the device updates the first parameter data in the deep learning network through automatic simulation until the deep learning network is successfully optimized; the image on the picture can be comprehensively, stably, accurately, quickly and in real time identified through the optimized and successful deep learning network; determining the category of the image on the picture and the coordinate data of the image on the picture; images of the same category and different scales on the image can be well identified; the method can be applied to the situation of identifying the images on the pictures on a large scale, liberates manual labor force and improves the efficiency of identifying the images on the pictures.

Drawings

Fig. 1 is a flowchart of a method for performing image recognition based on a deep learning network according to an embodiment of the present invention;

2-1 and 2-2 are flowcharts of a method for performing image recognition based on a deep learning network according to a second embodiment of the present invention;

fig. 3 is a flowchart of a method for performing image recognition based on a deep learning network according to a third embodiment of the present invention.

Detailed Description

The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the drawings in the embodiments of the present invention, and it is obvious that the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. All other embodiments, which can be derived by a person skilled in the art from the embodiments given herein without making any creative effort, shall fall within the protection scope of the present invention.

Example one

An embodiment of the present invention provides a method for performing image recognition based on a deep learning network, as shown in fig. 1, including the following steps:

step 100: the device transmits the first parameter data into a deep learning network, and acquires a picture sample library from a picture database;

step 101: when a picture is obtained from the picture sample library, dividing the picture into first preset data grids according to first preset data and calculating grid coordinate data of each grid; identifying the acquired picture by using a deep learning network comprising first parameter data according to first preset data, second preset data and third preset data to obtain a prediction frame data table, wherein the prediction frame data table comprises second preset data of prediction frame data corresponding to all grids respectively; sequentially acquiring grids and grid coordinate data corresponding to the grids from the first preset data grids, and executing a first operation on the acquired grids:

the first operation is: the device determines an image corresponding to the grid and actual coordinate data of the image corresponding to the grid according to the grid coordinate data and an image actual coordinate data table corresponding to the picture; sequentially acquiring the prediction frame data from the second preset data prediction frame data corresponding to the grid, and executing a second operation on the acquired prediction frame data: the second operation includes steps 101-01 to 101-03:

step 101-01: the device acquires all category probability data from the prediction frame data, selects the maximum category probability data from all category probability data, and determines the category data of the prediction frame data according to the maximum category probability data;

step 101-02: the device acquires an image corresponding to the grid and actual coordinate data of the image; sequentially acquiring the image and the actual coordinate data of the image from the image and the actual coordinate data of the image corresponding to the grid, and executing a third operation on the acquired image and the actual coordinate data of the image: and sequentially acquiring images from the same category image table corresponding to the prediction frame, and executing a fourth operation on the acquired images:

the third operation is: the device acquires the class data of the image from the actual coordinate data of the image, and correspondingly stores the prediction frame data and the actual data of the image and the image to a same class image table corresponding to the prediction frame when the acquired class data of the image is the same as the class data of the prediction frame data;

the fourth operation is: when the device acquires actual coordinate data of the images from the corresponding image tables of the same category, calculating the prediction frame data and the actual coordinate data to obtain a second confidence parameter; calculating the confidence corresponding to the image according to the first confidence parameter, the eleventh preset data, the twelfth preset data and the second confidence parameter in the prediction frame data, and correspondingly storing the confidence, the prediction frame data, the category data and the actual coordinate data of the image into a first list;

step 101-03: the device obtains the maximum confidence in the first list; judging whether the maximum confidence is greater than the first comparison data, if so, correspondingly storing prediction frame data corresponding to the maximum confidence and the actual coordinate data of the image, which are acquired from the first list, to a second list, and emptying the first list and the prediction frame corresponding to the image tables of the same category; otherwise, emptying the first list and the image list of the same category corresponding to the prediction frame;

step 102: when all pictures in the picture sample library are acquired, the device sequentially acquires the prediction frame data from the second list, and performs a fifth operation on the acquired prediction frame data:

optionally, the fifth operation comprises the steps of: the device calculates deviation data of the prediction frame data and the actual coordinate data according to the prediction frame data and the actual coordinate data, and stores the deviation data to a fifth list.

Step 103: the device obtains total deviation data according to all deviation data in the fifth list, compares the total deviation data with the preset deviation data, and executes step 104 when the total deviation data is greater than or equal to the preset deviation data; when the total deviation data is smaller than the preset deviation data, processing the picture to be identified by using a deep learning network containing first parameter data to obtain a prediction frame data table, and ending;

step 104: the device updates the first parameter data using a fourth preset algorithm based on the total deviation data and the preset deviation data, clears the second list and the fifth list, and returns to step 100.

Optionally, the grid data comprises grid abscissa, grid ordinate, grid length and grid width;

correspondingly, the prediction box data table comprises fourth data of prediction box data, and the prediction box data comprises a prediction box abscissa, a prediction box ordinate, a prediction box length, a prediction box width, a first confidence parameter and third preset data of category probability data;

correspondingly, the fourth data has a binding relationship with the first preset data and the second preset data, and further, the binding relationship is as follows: the fourth data is the product of the first preset data and the second preset data;

further correspondingly, the first preset data is in the format of data S.

Optionally, the third operation comprises the following steps D01 to D02:

step D01: the device acquires the category data of the image from the actual coordinate data of the image;

step D02: the device judges whether the category data of the image is the same as the category data of the prediction frame data, if so, the actual coordinate data and the prediction frame data of the image are stored in the same category image table of the prediction frame; otherwise, ending.

Optionally, the fourth operation comprises the following steps D11 to D12;

step D11: the device obtains the intersection area and the union area of the prediction frame data and the image according to the eleven preset algorithm; calculating the intersection area and the union area to obtain a second confidence parameter;

further, in step D11, the apparatus obtains an intersection area and a union area of the prediction frame data and the image according to the prediction frame data and the actual data of the image, specifically: and according to an eleventh preset algorithm, calculating the horizontal coordinate, the vertical coordinate, the length and the width of the prediction frame data and the actual horizontal coordinate, the actual vertical coordinate, the actual length and the actual width of the actual data of the image to obtain the intersection area and the union area of the prediction frame corresponding to the first current prediction frame data and the second current image.

Further, in step D11, the intersection area and the union area are calculated to obtain a second confidence parameter, which specifically is: the device calculates the intersection area and the union area according to a first preset algorithm to obtain a second confidence parameter.

Step D12: the device acquires a first confidence parameter in the prediction frame data, and determines the confidence of the prediction frame data according to the first confidence parameter, the second confidence parameter, the fifth preset data and the sixth preset data; correspondingly storing the confidence coefficient of the prediction frame data, the category data of the prediction frame data and the actual data of the image into a first list;

further, step D12 is specifically: the device acquires a first confidence parameter in the prediction frame data, determines the category of the first confidence parameter, determines the confidence coefficient of the prediction frame data according to fifth preset data and a second confidence parameter when the first confidence parameter is fifth preset data, and correspondingly stores the confidence coefficient of the prediction frame data, the category data and the actual data confidence coefficient of the image to a first list; when the first confidence parameter is sixth preset data, determining the confidence coefficient of the prediction frame data according to the second confidence parameter, and correspondingly storing the confidence coefficient of the prediction frame data, the category data and the actual data of the image into a first list;

further, step D12 specifically includes: the device acquires a first confidence parameter in the prediction frame data, determines the category of the first confidence parameter, when the first confidence parameter is fifth preset data, calculates the fifth preset data and the second confidence parameter according to a second preset algorithm to obtain the confidence coefficient of the prediction frame data, and correspondingly stores the confidence coefficient of the prediction frame data, the category data and actual data of an image to a first list; and when the first confidence parameter is sixth preset data, recording the second confidence parameter as the confidence coefficient of the prediction frame data, and correspondingly storing the confidence coefficient of the prediction frame data, the category data and the actual data of the image into the first list.

Further, in step D11, the intersection area and the union area are calculated to obtain a second confidence parameter, which specifically is: the device calculates the intersection area and the union area according to a first preset algorithm to obtain a second confidence parameter;

further, in step D11, the intersection area and the union area are calculated to obtain a second confidence parameter, which specifically is: the device carries out ratio operation on the intersection area and the union area according to a first preset algorithm to obtain a second confidence parameter;

further, in step 101-03, the prediction frame data and the actual coordinate data of the image, which are acquired from the first list and correspond to the maximum confidence level, are correspondingly stored in the second list, specifically: acquiring the prediction frame data corresponding to the maximum confidence coefficient and the actual coordinate data of the image from the first list, and correspondingly storing the maximum confidence coefficient, the prediction frame data and the actual coordinate data of the image into a second list; the device obtains a fourth list from the second list;

the apparatus obtains a fourth list from the second list, comprising the steps of:

step D20: the device classifies the prediction frame data in the second list, and selects one class from the obtained classes as a first current class;

step D21: the device acquires the prediction frame data corresponding to the first current category, the actual data of the image and the confidence coefficient of the prediction frame data from the second list and stores the prediction frame data, the actual data of the image and the confidence coefficient of the prediction frame data in a third list;

step D22: the device acquires the maximum confidence from the third list, marks the prediction frame data corresponding to the maximum confidence, and marks the prediction frame data corresponding to the maximum confidence as the first marked prediction frame data;

step D23: the apparatus selects one of the prediction box data other than the first marked prediction box data from the third list as a second current prediction box data;

step D24: the device acquires actual data corresponding to the second current prediction frame data from the third list; obtaining intersection area and union area of the prediction frame corresponding to the first mark prediction frame data and the prediction frame corresponding to the second current prediction frame data according to the prediction frame data corresponding to the first mark prediction frame data and the prediction frame data of the second current prediction frame; calculating the intersection area and the union area to obtain a third confidence parameter;

step D25: the apparatus compares the third confidence parameter with the third comparison data, and performs step D26 when the third confidence parameter is greater than or equal to the third comparison data; when the third confidence parameter is less than the third comparison data, marking a second current prediction box in a third list, and executing step D26;

step D26: the apparatus determines whether there is prediction box data that is not regarded as the second current prediction box data in the prediction box data other than the marked prediction box data in the third list, and if so, performs step D27; otherwise, correspondingly storing the prediction box marked in the third list and the corresponding confidence coefficient, actual data, prediction box data and category data into a fourth list, emptying the third list, and executing the step D28;

step D27: the apparatus acquires the next prediction box data from the prediction boxes in the third list other than the first-marked prediction box data as the second current prediction box data, and returns to step D24.

Step D28: the device determines whether there is any category in the second list that is not considered as the first current category, if so, executes step D29; otherwise, emptying the second list, and executing a fifth operation on each prediction box data in the fourth list;

step D29: the apparatus selects the next category from the categories in the second list as the first current category and returns to step D21.

Optionally, in step 103, processing the picture to be recognized by using a deep learning network including first parameter data to obtain a prediction frame data table, including the following steps:

step T1: when the picture to be identified is obtained, the picture is divided into grids of first preset data by the device according to the first preset data; identifying the picture to be identified by using a deep learning network comprising updated first parameter data according to the first preset data, the second preset data and the third preset data to obtain a prediction frame data table; the prediction frame data table comprises second preset data prediction frame data respectively corresponding to all grids; acquiring grids from the prediction frame data table in sequence, and performing an eleventh operation on the grids:

the eleventh operation is: acquiring prediction frame data from second preset data prediction frame data corresponding to the grids in sequence, and performing twelfth operation on the prediction frame data;

the twelfth operation is: the device acquires a first confidence parameter from the prediction frame data, and marks the prediction frame data in a prediction frame data table when the first confidence parameter is fifth preset data;

step T2: the device outputs the prediction frame data of all marks in the prediction frame data table;

further, in the twelfth operation, the prediction box data is marked in the prediction box data table, and replaced with: the device saves the prediction box data to a twelfth list; acquiring the prediction frame data from the twelfth list in sequence, and executing a thirteenth operation on the prediction frame data;

the thirteenth operation is that when one piece of prediction frame data is obtained from the twelfth list, the device obtains all kinds of probability data from the prediction frame data, selects the largest kind of probability data from the obtained all kinds of probability data, determines the kind data of the prediction frame data according to the largest kind of probability data, and correspondingly stores the kind data and the prediction frame data in the twelfth list; classifying all the prediction frame data in the twelve lists according to all the category data in the twelve lists; acquiring any category from the twelve lists, executing a fourteenth operation, and emptying the twelfth list when all the categories are acquired from the twelve lists;

fourteen operations are as follows: the device acquires the prediction frame data corresponding to the category from the twelfth list and stores the prediction frame data in the thirteenth list; acquiring the largest category probability data of all category probability data corresponding to the category from the prediction frame data in the thirteenth list; marking the prediction box data corresponding to the maximum category probability data in a thirteenth list; according to an eleventh preset algorithm, calculating the prediction frame data corresponding to the maximum category probability data and other prediction frame data in the thirteenth list respectively to obtain sixth confidence parameters corresponding to other prediction frame data in the thirteenth list respectively; performing a fifteenth operation on each sixth confidence parameter; the fifteenth operation is: the apparatus compares the magnitude of the sixth confidence parameter and the sixth comparison data, and marks the prediction box data in the thirteenth list when the sixth confidence parameter is smaller than the thirteenth confidence data; correspondingly marking the prediction frame data of all marks in the thirteenth list in a prediction frame data table; emptying the thirteenth list;

further, according to an eleventh preset algorithm, the prediction frame data corresponding to the maximum category probability data and other prediction frame data in the thirteenth list are respectively calculated to obtain sixth confidence parameters respectively corresponding to other prediction frame data in the thirteenth list, specifically:

the device acquires prediction frame data corresponding to the maximum category probability data from the thirteenth list and records the prediction frame data as marked prediction frame data; when a prediction box data except the marker prediction box data is obtained from the thirteenth list, the device calculates the marker prediction box data and the second current prediction box data according to an eleventh preset algorithm to obtain an intersection area and a union area, and calculates the intersection area and the union area according to the first preset algorithm to obtain a sixth confidence parameter.

Example two

The second embodiment of the invention provides a method for image recognition based on a deep learning network; as shown in fig. 2-1 and 2-2, comprising the steps of:

prior to this embodiment, the user sets the following data in advance in the device:

1. the picture sample library is a part of the picture database; in the embodiment, 10 pictures are selected from the picture database;

2. actual data of the image on each picture, actual data (an actual abscissa of the central point, a calibration ordinate of the central point, an actual length, an actual width, category data), wherein the actual abscissa, the actual length and the actual width in the actual data are pixel coordinates;

3. wherein 3.1 is the first preset data (7X7 grids, which is 7X7 in this embodiment); 3.2 second preset data (2 in the present embodiment) indicating the number of prediction boxes that can be recognized in each mesh; 3.3, third preset data (2 in the present embodiment) indicating the number of categories of images on the picture input into the deep learning network, wherein the first category is origin, and the second category is bdm;

4. first comparison data 0.142857143 and third comparison data 0.8;

5. presetting deviation data 3.5;

6. the first parameter data is the parameter data required to be transmitted by the deep learning network; the initial time is the preset data transmitted by the user; then the device updates the first parameter data according to the actual situation and then transmits the first parameter data into the deep learning network again; for example, the first parameter data is matrix data, including bias/weight;

step 200: the device transmits the first parameter data into a deep learning network, and acquires a picture sample library from a picture database;

optionally, the step specifically includes: the device randomly obtains a picture sample library from a picture database;

for example, the device randomly acquires 10 pictures from a picture database as a picture sample library;

step 201: the device acquires a picture from a picture sample library as a first current picture;

optionally, this step is preceded by: the device preprocesses a sample picture in a picture database, puts the picture generated by preprocessing into the picture database of the sample picture and stores actual data of an image on the picture generated by preprocessing;

the pretreatment in this step is aimed at two purposes: firstly, the sample capacity in the deep learning process is increased; secondly, images in the pictures are better identified; the preprocessing comprises smoothing and/or median filtering and/or geometric transformation and/or normalization and/or smoothing and/or restoration and/or enhancement and/or cropping and/or scaling and/or flipping and/or shifting and/or adding noise;

step 202: the device acquires actual data of all images in a first current picture and stores the actual data into an image actual coordinate data table;

for example, there are 8 images on the first current picture, and the actual data of all the images on the first current picture is shown in the following table 11:

375	279	16	16	orifice
					357	108	16	16	orifice
502.5	180	17	18	orifice
					279	363	18	18	orifice
458	322.5	18	17	orifice
					508.5	484.5	19	19	orifice
639	234	20	18	orifice
					768	211	18	18	orifice

step 203: the device divides a first current picture into 7X7 grids of first preset data to obtain a grid coordinate data table comprising 49 grid coordinate data of the first preset data, wherein the grid coordinate data comprise grid horizontal coordinates, grid vertical coordinates, grid length and grid width;

before the step, a user presets first preset data in the device; the first preset data indicates the number of the grids into which the picture is divided; for example, the first preset data is 7X7, that is, the image is divided into 7 mesh lengths equally in the transverse direction and 7 mesh widths equally in the longitudinal direction; dividing the first current picture into 7X7 grids by the user, wherein each grid corresponds to a grid coordinate data, and one grid coordinate data comprises a central coordinate (a grid horizontal coordinate and a grid vertical coordinate), a grid length and a grid width;

for example, the data format of each grid coordinate data is: (grid abscissa of grid center, grid ordinate of grid center, grid length, grid width);

the 7X7 grid coordinate data on the first current picture are shown in table 12 below:

step 204: the device identifies the first current picture by using a deep learning network comprising first parameter data to obtain fourth data (7X7X2) of prediction box data, saves the fourth data (7X7X2) of the prediction box data to a prediction box data table, and the prediction box data comprises 2 category probability data of a prediction box abscissa, a prediction box ordinate, a prediction box length, a prediction box width, a first confidence parameter and third preset data;

optionally, each prediction frame data includes (4+1+ third preset data 2) sub data, that is, a midpoint coordinate of the prediction frame, a prediction frame height and a prediction frame width of the prediction frame, and category probability data of each category of the third preset data C to which the target image in the prediction frame belongs;

the prediction frame data is rectangular frame data containing each image recognized by the device; the horizontal coordinate of the prediction frame is the horizontal coordinate of the middle point of the rectangular frame; the vertical coordinate of the prediction frame is the vertical coordinate of the middle point of the rectangular frame; predicting the frame length to be the length of the rectangular frame; predicting the frame width to be the width of the rectangular frame;

before this embodiment, the user sets the second preset data 2 and the third preset data 2 in the device in advance; the second preset data represents the number of prediction boxes that can be identified in each mesh; the third preset data represents the total number of categories contained in the images input to the pictures in the deep learning network; 4, four subdata including a prediction frame abscissa and a prediction frame ordinate of the midpoint of the prediction frame, a prediction frame length and a prediction frame width of the prediction frame are represented, and pixel coordinates are used for all coordinates, lengths and widths on the image; 1 denotes the sub-data of the first confidence parameter for each prediction box; the prediction box data may be displayed as parenthesis (prediction box abscissa of midpoint of prediction box, prediction box ordinate of midpoint, prediction box length, prediction box width, first confidence parameter, third preset data category probability data); the first confidence parameter represents whether a target object falls into the grid corresponding to the prediction box, if the target object is 1, and if the target object is not 0;

for example, in the present embodiment, the image categories include two categories, i.e., orifice and bdm, and the prediction frame data is shown in the following table:

the first current picture derives 7X2 predictor data and stores it in a predictor data table, which is shown in the following figure:

step 205: the apparatus acquires one grid 5X6 of the first preset data 7X7 grids as a first current grid;

step 206: the device acquires 2 pieces of second preset data corresponding to 5x6 of the first current grid from the prediction frame data table and stores the second preset data into the grid corresponding prediction frame table;

step 207: the device acquires a prediction frame data from a grid corresponding prediction frame table as a first current prediction frame data;

for example, the number of the prediction boxes corresponding to the grid is 2, and one prediction box data (365, 280, 15, 18, 1, 69.32850645, 30.67149355) is arbitrarily selected from the second preset data as the first current prediction box data;

step 208: the device acquires first current prediction frame data from a grid corresponding prediction frame table; acquiring 2 category probability data of third preset data from the first current prediction frame data; selecting the maximum class probability data from the 2 class probability data of the third preset data, and determining the class data of the first current prediction frame data according to the maximum class probability data;

for example, the apparatus obtains first current predictor data from a predictor block data table (365, 280, 15, 18, 1, 69.32850645, 30.67149355); obtaining category probability data corresponding to the first current prediction box data from the prediction box data (69.32850645, 30.67149355); acquiring 2 types of probability data of third preset data corresponding to the first current prediction frame data from the prediction frame data; selecting the largest category probability data (69.32850645) from the third preset data 2 category probability data (69.32850645, 30.67149355), and acquiring the category data origin of the first current prediction frame data according to the largest category probability data (69.32850645);

step 209: the device determines the image and the actual data in the first current grid according to the image actual coordinate data table and the grid coordinate data of the first current grid; saving the image in the first current grid and the actual data of the image to the grid corresponding image table, and executing step 209-01;

step 209-01: the device acquires an image from the grid corresponding image table as a first current image;

step 209-02: the device acquires actual coordinate data of a first current image, and acquires category data of the first current image from the actual coordinate data of the first current image;

step 209-03: the device judges whether the category data of the first current image is the same as the category data of the first current prediction frame data, if so, the actual coordinate data of the first current image and the first current prediction frame data are stored in a same category image table of the prediction frame, and step 209-04 is executed, otherwise, step 209-04 is executed;

step 209-04: the device judges whether the next image as the first current image can be obtained from the grid corresponding image table, if yes, the step is returned to 209-05; otherwise, executing step 209-06;

step 209-05: the device acquires the next image from the grid corresponding image table as the first current image and returns to the step 209-02;

step 209-06: the apparatus obtains an image from the same category image table of the prediction frame as a second current image, and performs step 210;

step 210: the device acquires actual data of a second current image from the same-class image table of the prediction frame; obtaining the intersection area and the union area of the first current prediction frame data and the second current image according to the first current prediction frame data and the actual data of the second current image; calculating the intersection area and the union area to obtain a second confidence parameter;

optionally, the step specifically includes: the device acquires actual data of a second current image from the image actual coordinate data table; calculating the horizontal coordinate, the vertical coordinate, the length and the width of the prediction frame of the first current prediction frame data and the actual horizontal coordinate, the actual vertical coordinate, the actual length and the actual width of the actual data of the second current image according to an eleventh preset algorithm to obtain the intersection area and the union area of the prediction frame corresponding to the first current prediction frame data and the second current image; calculating the intersection area and the union area by using a first preset algorithm to obtain a second confidence parameter;

optionally, the step is more specifically: the device acquires actual data of a second current image from the image actual coordinate data table; calculating the horizontal coordinate, the vertical coordinate, the length and the width of the prediction frame of the first current prediction frame data and the actual horizontal coordinate, the actual vertical coordinate, the actual length and the actual width of the actual data of the second current image according to an eleventh preset algorithm to obtain the intersection area and the union area of the prediction frame corresponding to the first current prediction frame data and the second current image; performing ratio operation on the intersection area and the union area to obtain a second confidence parameter;

for example, the apparatus acquires actual data (375, 279, 16, 16, origin) of the second current image from the image actual coordinate data table; deriving an intersection area 88 and a union area 438 of the first current prediction box data and the second current image from the first current prediction box data (365, 280, 15, 18, 1, 69.32850645, 30.67149355) and actual data (375, 279, 16, origin) of the second current image; performing a ratio operation on the intersection area 88 and the union area 438 to obtain a second confidence parameter 0.200913242;

step 211: the device acquires a first confidence parameter in the first current prediction frame data, and determines the confidence of the first current prediction frame data according to the first confidence parameter, the second confidence parameter, the fifth preset data and the sixth preset data; correspondingly storing the confidence coefficient of the first current prediction frame data, the prediction frame data and the category data and the actual data of the second current image into a first list;

optionally, the step specifically includes: the device acquires a first confidence parameter in first current prediction frame data, determines the type of the first confidence parameter, determines the confidence coefficient of the first current prediction frame data according to fifth preset data and a second confidence parameter when the first confidence parameter is fifth preset data, stores the confidence coefficient of the first current prediction frame data, the prediction frame data and the type and actual data of a second current image into a first list, and executes step 212; when the first confidence parameter is sixth preset data, determining the confidence of the first current prediction frame data according to the second confidence parameter, correspondingly storing the confidence of the first current prediction frame data, the prediction frame data and the type and the actual data of the second current image in a first list, and executing step 212;

optionally, the step is more specifically: the device acquires a first confidence parameter in first current prediction frame data, determines the type of the first confidence parameter, when the first confidence parameter is fifth preset data, calculates the fifth preset data and the second confidence parameter according to a second preset algorithm to obtain the confidence coefficient of the first current prediction frame data, correspondingly stores the confidence coefficient of the first current prediction frame data, the prediction frame data and the type and actual data of a second current image into a first list to a first list, and executes step 212; when the first confidence parameter is the sixth preset data, the second confidence parameter is the confidence of the first current prediction frame data, the confidence of the corresponding first current prediction frame data, the prediction frame data and the type and the actual data of the second current image are stored in the first list to the first list, and step 212 is executed;

for example, the apparatus obtains a first confidence parameter in first current prediction box data (365, 280, 15, 18, 1, 69.32850645, 30.67149355), determines a category of the first confidence parameter, when the first confidence parameter is fifth preset data 1, calculates the fifth preset data and the second confidence parameter 0.200913242 according to a second preset algorithm to obtain a confidence 0.149980763498417 of the first current prediction box data, correspondingly stores the confidence 0.149980763498417 of the first current prediction box data, the prediction box data (365, 280, 15, 18, 1, 69.32850645, 30.67149355), the orifice of the category data and the actual data (375, 279, 16, orifice) of the second current image in a first list to a first list, and executes step 212; when the first confidence parameter is the sixth preset data, the second confidence parameter is the confidence of the first current prediction frame data, the prediction frame data and the type and the actual data of the second current image are correspondingly stored in the first list, and the step 212 is executed;

in this embodiment, the confidence degree represents the confidence degree of the image recognized by the prediction box;

step 212: the apparatus determines whether there is any image in the same category image table of the prediction frame that is not used as the second current image, if yes, performs step 213; otherwise, go to step 214;

step 213: the device obtains the next image from the same-class image table of the prediction frame as the second current image, and returns to step 210;

step 214: the device empties the image tables of the same category of the prediction frame, and obtains the maximum confidence coefficient from the first list to be recorded as a first current confidence coefficient; comparing the first current confidence coefficient with the first comparison data, and when the first current confidence coefficient is larger than the first comparison data, acquiring actual data of the image corresponding to the first current confidence coefficient and category data of the first current prediction frame data from the first list; correspondingly saving the first current confidence, the actual data of the image, the first current prediction frame data and the category data of the first current prediction frame data to a second list, and executing step 215; when the confidence of the first current prediction box data is less than or equal to the first comparison data, performing step 215;

for example, the device obtains the maximum confidence from the first list and records the maximum confidence as the first current confidence; comparing the first current confidence with the first comparison data 0.142857142857143, and when the first current confidence is greater than the first comparison data, acquiring the actual data of the image corresponding to the first current confidence, the first current prediction frame data, and the category data of the first current prediction frame data from the first list; correspondingly saving the first current confidence, the actual data of the image, the first current prediction frame data and the category data of the first current prediction frame data to a second list, and executing step 215; when the confidence of the first current prediction box data is less than or equal to the first comparison data, performing step 215;

step 215: the device empties the first list; judging whether 2 prediction frames corresponding to the first current grid have prediction frames which are not used as data of the first current prediction frame, if so, executing step 216; otherwise, executing step 217;

step 216: the device obtains the next prediction frame from the 2 prediction frames corresponding to the first current grid as the data of the first current prediction frame, and returns to step 208;

step 217: the device classifies the prediction frame data in the second list according to the class data in the second list; selecting a category as a first current category;

for example, all 8 images on the first current picture are selected as the origin, so that the grid corresponding prediction frame table only has one category origin;

step 218: the device acquires the prediction frame data corresponding to the first current category, the actual data of the image and the confidence coefficient of the prediction frame data from the second list and stores the prediction frame data, the actual data of the image and the confidence coefficient of the prediction frame data in a third list; acquiring the maximum confidence from the third list, marking the prediction frame data corresponding to the maximum confidence, and recording the prediction frame data corresponding to the maximum confidence as marked prediction frame data;

step 219: the apparatus selects one of the prediction box data other than the marked prediction box data from the third list as a second current prediction box data;

step 220: the device acquires actual data corresponding to the second current prediction frame data from the third list; obtaining intersection area and union area of the prediction frame corresponding to the mark prediction frame data and the prediction frame corresponding to the second current prediction frame data according to the prediction frame data corresponding to the mark prediction frame data and the prediction frame data of the second current prediction frame; calculating the intersection area and the union area to obtain a third confidence parameter;

step 221: the device compares the third confidence parameter with the third comparison data, and executes step 222 when the third confidence parameter is greater than or equal to the third comparison data; when the third confidence parameter is less than the third comparison data, marking a second current prediction box in a third list, and executing step 222;

for example, the apparatus compares the third confidence parameter with the third comparison data by 0.8, and performs step 222 when the third confidence parameter is greater than or equal to the third comparison data; when the third confidence parameter 0 is less than the third comparison data 0.8, marking a second current prediction box in the third list, and executing step 222;

step 222: the apparatus determines whether there is prediction box data that is not regarded as the second current prediction box data in the prediction box data other than the marked prediction box data in the third list, if so, performs step 223; otherwise, correspondingly saving the prediction frame data marked in the third list and the corresponding actual data of the image into a fourth list, emptying the third list, and executing step 224;

step 223: the apparatus acquires next prediction box data from the prediction box data in the third list except the marked prediction box data as second current prediction box data, and returns to step 220;

step 224: the device determines whether there is a category that is not considered as the first current category in the categories in the second list, if so, performs step 225; otherwise, emptying the second list, and performing step 226;

step 225: the device obtains the next category from the categories in the second list as the first current category, and returns to step 218;

step 226: the apparatus determines whether there is any mesh that is not used as the first current mesh among the 7 × 7 meshes of the first preset data, if yes, performs step 227; otherwise, go to step 228;

step 227: the apparatus obtains the next mesh from the 7x7 meshes as the first current mesh, and returns to step 206;

step 228: the device determines whether there is any picture in the picture sample library that is not taken as the first current picture, if so, performs step 229; otherwise, go to step 231;

step 229: the device obtains the next picture from the picture sample library as the first current picture, and returns to step 202;

step 231: the device acquires a prediction frame data from the fourth list as a third current prediction frame data;

step 232: the apparatus acquires actual data of an image corresponding to the third current prediction frame data from the fourth list; calculating the third current prediction frame data and actual data corresponding to the third current prediction frame data according to a third preset algorithm to obtain deviation data of the third current prediction frame data; storing deviation data of the third current prediction frame data to a fifth list;

step 233: the apparatus determines whether there is any prediction box data in the fourth list that is not regarded as the third current prediction box data, if yes, performs step 234; otherwise, clear the fourth list, go to step 235;

step 234: the apparatus obtains the next prediction box data from the fourth list as the third current prediction box data, and returns to step 232;

step 235: the apparatus obtains total deviation data according to the deviation data in the fifth list, compares the total deviation data with the preset deviation data, clears the fifth list when the total deviation data is greater than or equal to the preset deviation data, and executes step 236; when the total deviation data is smaller than the preset deviation data, processing the picture to be identified by using a deep learning network containing first parameter data to obtain a prediction frame data table, and ending;

step 236: the device updates the first parameter data by using a fourth preset algorithm according to the total deviation data and the preset deviation data, and returns to the step 200;

for example, the fourth preset algorithm is a back propagation method;

the total loss value was 213.6

The preset offset data is 3.0.

EXAMPLE III

The third embodiment of the invention provides a method for image recognition based on a deep learning network; the deep learning network used in the embodiment is the optimized and successful deep learning network in the second embodiment; as shown in fig. 3, the method comprises the following steps:

step 301: the device receives a picture to be identified, which is transmitted by a user;

step 302: the device uses a deep learning network comprising first parameter data to identify a picture to be identified to obtain 7x7x2 prediction frame data and corresponding prediction frame data respectively, and stores the prediction frame data and the prediction frame data thereof to a prediction frame data table; the prediction frame data comprises 2 category probability data of a prediction frame data abscissa and a prediction frame data ordinate of a midpoint of the prediction frame data, a prediction frame data length, a prediction frame data width, a first confidence parameter and third preset data;

in a second embodiment before this embodiment, a user sets first preset data in the device in advance; the first preset data indicates the number of meshes into which the image is divided; the first preset data is generally 7x7, namely, the image is divided into S grid lengths in the transverse direction and S grid widths in the longitudinal direction;

optionally, each of the predicted frame data includes (4+1+ C) sub-data, that is, a midpoint coordinate (midpoint abscissa and midpoint ordinate) of the predicted frame data, a predicted frame data height and a predicted frame data width of the predicted frame data, and category probability data of the target image in the predicted frame data belonging to each of the C categories; presetting data B and data C in the device by a user; data B represents the number of prediction box data that can be identified in each mesh; data C represents the image category of the image on the picture input in the deep learning network; 4, representing four subdata of a midpoint coordinate of the prediction frame data, a prediction frame data length of the prediction frame data and a prediction frame data width, wherein all coordinates, lengths and widths on the image use pixel coordinates; 1 denotes the sub-data of the first preset parameter of each prediction box data; the prediction box data may be displayed as: (center abscissa, center ordinate, prediction box data length, prediction box data width, first confidence parameter, C class probability data);

for example, when C is 2, the prediction box data is as shown in the following table:

step 303: the apparatus acquires one mesh from the first preset data 7X7 meshes as a first current mesh;

step 304: the device acquires second preset data 2 prediction frame data corresponding to the first current grid from the prediction frame data table and stores the second preset data in the prediction frame table corresponding to the grid;

step 305: the device acquires a prediction frame data from the grid corresponding prediction frame table as a fifth current prediction frame data;

step 306: the device acquires a first confidence parameter from the fifth current prediction frame data, determines the category of the first confidence parameter, stores the fifth current prediction frame data to a twelfth list when the first confidence parameter is fifth preset data, and executes step 307; when the first confidence parameter is the sixth preset data, executing step 307;

step 307: the device judges whether the next prediction frame data as the fifth current prediction frame data can be obtained from the grid corresponding prediction frame table, if so, the next prediction frame data is obtained from the grid corresponding prediction frame table as the fifth current prediction frame data, and the step 306 is returned; otherwise, go to step 308;

step 308: the device acquires a prediction frame data from the twelfth list as a sixth current prediction frame data;

step 309: the device acquires 2 category probability data of third preset data from sixth current prediction frame data acquired from a twelfth list; selecting the maximum class probability data from the 2 class probability data of the third preset data, and determining the class data of the sixth current prediction frame data according to the maximum class probability data; correspondingly storing the category data of the sixth current prediction frame data and the sixth current prediction frame data into a twelfth list;

step 310: the device determines whether the next prediction frame data serving as sixth current prediction frame data can be acquired from the twelfth list, if so, the device acquires the next prediction frame data serving as sixth current prediction frame data from the twelfth list, and returns to step 309; otherwise, go to step 311;

step 311: the device classifies the prediction frame data in the twelfth list according to the category data in the twelfth list; selecting a category as a second current category;

step 312: the device acquires the prediction frame data corresponding to the second current category from the twelfth list and stores the prediction frame data into the thirteenth list;

step 313: the device acquires the largest category probability data of all category probability data corresponding to the second current category from the prediction frame data in the thirteenth list; marking the prediction box data corresponding to the maximum category probability data in a thirteenth list;

step 314: the apparatus records, in the thirteenth column, the prediction box data corresponding to the largest class probability data as second labeled prediction box data;

step 315: the apparatus selects one of the predicted frame data other than the second marked predicted frame data from the thirteenth list as second current predicted frame data;

step 316: the device calculates a sixth confidence parameter for the second mark prediction frame data and the second current prediction frame data according to an eleventh preset algorithm;

step 317: the apparatus compares the magnitude of the sixth confidence parameter and the sixth comparison data, and performs step 318 when the sixth confidence parameter is greater than or equal to the sixth comparison data; when the sixth confidence parameter is less than the thirteenth confidence data, marking the second current prediction box data in the thirteenth list, and executing step 318;

step 318: the apparatus determines whether there is prediction box data that is not regarded as second current prediction box data in the prediction box data other than the marked prediction box data in the thirteenth list, if so, performs step 319; otherwise, correspondingly marking the prediction box data of all marks in the thirteenth list in the prediction box data table, and executing step 320;

step 319: the apparatus acquires the next prediction box data from the prediction box data in the thirteenth list excluding the second marked prediction box data as the second current prediction box data, and returns to step 316;

step 320: the device empties the thirteenth list; judging whether the categories in the twelfth list have categories which are not used as the second current category, if so, executing step 321; otherwise go to step 322;

step 321: the device obtains the next category from the categories in the twelfth list as the second current category, and returns to step 312;

step 322: the apparatus empties the twelfth list; judging whether grids which are not used as the second current grid exist in the 7x7 grids of the first preset data, if so, acquiring the next grid from the 7x7 grids of the first preset data as the second current grid, and returning to the step 304; otherwise, outputting the prediction frame data marked in the prediction frame data table, and ending.

Example four

The fourth embodiment provides a device for image recognition based on a deep learning network, which comprises an incoming module, an acquisition module, an execution module, a deviation comparison module, a module to be recognized and an update parameter module;

the acquisition module is also used for acquiring a picture from the picture sample library, dividing the picture into first preset data grids according to the first preset data and calculating grid coordinate data of each grid; identifying the acquired picture by using a deep learning network comprising first parameter data according to first preset data, second preset data and third preset data to obtain a prediction frame data table, wherein the prediction frame data table comprises second preset data of prediction frame data corresponding to all grids respectively; acquiring grids and grid coordinate data of the grids from the first preset data grids in sequence;

an execution module, configured to perform a first operation on the mesh acquired by the acquisition module:

the execution module comprises a first execution operation unit;

a second execution operation unit for second-operating the acquired prediction box data of the determination module unit:

the first determining unit is used for acquiring all category probability data from the acquired prediction frame data of the determining module unit, selecting the largest category probability data from all category probability data, and determining the category data of the prediction frame data according to the largest category probability data; acquiring an image corresponding to a grid and actual coordinate data of the image; acquiring images and actual coordinate data of the images from the images and the actual coordinate data of the images corresponding to the grids in sequence, and triggering a third execution operation unit;

a third performing operation unit configured to perform a third operation on the image and the actual coordinate data of the image acquired by the first determining unit:

a third performing unit configured to acquire category data of the image from the actual coordinate data of the image acquired by the first determining unit, and when the acquired category data of the image is the same as the category data of the prediction frame data, correspondingly store the prediction frame data and the image and the actual data of the image into a prediction frame corresponding same category image table;

the second acquisition unit is used for sequentially acquiring images from the same-class image table corresponding to the prediction frame obtained by the third execution operation unit;

the fourth execution operation unit is used for calculating the prediction frame data determined by the determination module unit and the actual coordinate data of the image acquired by the second acquisition unit to obtain a second confidence parameter; calculating the confidence corresponding to the image according to the first confidence parameter, the eleventh preset data, the twelfth preset data and the second confidence parameter in the prediction frame data, and correspondingly storing the confidence, the prediction frame data, the category data and the actual coordinate data of the image into a first list;

the acquisition module is further used for acquiring the prediction frame data and the actual coordinate data of the corresponding image from the second list when all the pictures in the picture sample library are acquired, and triggering a fifth execution operation unit;

a fifth execution operation unit, configured to calculate deviation data of the prediction frame data and the actual coordinate data according to the prediction frame data and the actual coordinate data, and store the deviation data in a fifth list;

the deviation comparison module is used for obtaining total deviation data according to all deviation data in the fifth list obtained by the fifth execution operation unit and comparing the total deviation data with the preset deviation data;

the processing to-be-identified module is used for successfully optimizing the deep learning network when the total deviation data obtained by the deviation comparison module is smaller than the preset deviation data, processing the to-be-identified picture by using the deep learning network containing the first parameter data to obtain a prediction frame data table, and ending;

and the parameter updating module is used for updating the first parameter data by using a fourth preset algorithm according to the total deviation data and the preset deviation data and emptying the second list and the fifth list when the total deviation data obtained by the deviation comparing module is smaller than the preset deviation data.

accordingly, the predicted frame data is rectangular frame data of a rectangular frame framing the image recognized by the apparatus; the horizontal coordinate of the prediction frame is the horizontal coordinate of the middle point of the rectangular frame; the ordinate of the prediction frame is the ordinate of the midpoint of the rectangular frame; predicting the length of the frame to be the length of the rectangular frame; predicting the width of the frame to be the width of the rectangular frame;

correspondingly, the fourth data has a binding relation with the first preset data and the second preset data;

further, the binding relationship is as follows: the fourth data is the product of the first preset data and the second preset data;

further correspondingly, the first preset data is the product of the data S and the data S.

Optionally, the save acquisition unit comprises an acquisition as subunit

Correspondingly, the acquisition as a subunit is used for acquiring one image from all the images corresponding to the grid as an image;

accordingly, an acquisition subunit configured to acquire actual coordinate data of the image acquired as the subunit, and acquire category data of the image from the actual coordinate data of the image acquired as the subunit;

correspondingly, the first judging subunit is used for judging whether the class data of the image as the subunit is acquired to be the same as the class data of the prediction frame data, and if so, the saving subunit is triggered;

correspondingly, the storage subunit is used for storing the actual coordinate data and the prediction frame data of the image as the subunit to the same-class image table of the prediction frame, and triggering the second judgment subunit, otherwise, triggering the second judgment subunit;

correspondingly, the second judging subunit is used for judging whether a next image can be acquired from all the images corresponding to the grid as an image, if so, acquiring the next image from all the images corresponding to the grid as the image, and triggering the acquiring subunit; and if not, acquiring images from the same category image tables corresponding to the prediction frames in sequence, and triggering a fourth operation execution unit.

Optionally, the fourth performing operation unit comprises a first calculating subunit and a first saving subunit;

correspondingly, the first calculating subunit is used for obtaining intersection area and union area of the prediction frame data and the image according to the eleven preset algorithm for the prediction frame data and the actual data of the image acquired by the second acquiring unit; calculating the intersection area and the union area to obtain a second confidence parameter;

correspondingly, the first saving subunit is used for acquiring a first confidence parameter in the prediction frame data and determining the confidence of the prediction frame data according to the first confidence parameter, the second confidence parameter of the first calculating subunit, the fifth preset data and the sixth preset data; and correspondingly storing the confidence coefficient of the prediction frame data, the category data of the prediction frame data and the actual data of the image into the first list.

Further, the first saving subunit is specifically configured to acquire a first confidence parameter in the prediction frame data, determine a category of the first confidence parameter, determine, when the first confidence parameter is fifth preset data, a confidence degree of the prediction frame data according to the fifth preset data and the second confidence parameter, and correspondingly save the confidence degree of the prediction frame data, the category data, and the confidence degree of the actual data of the image in the first list; when the first confidence parameter is sixth preset data, determining the confidence coefficient of the prediction frame data according to the second confidence parameter, and correspondingly storing the confidence coefficient of the prediction frame data, the category data and the actual data of the image into a first list;

furthermore, the first saving subunit is specifically configured to acquire a first confidence parameter in the prediction frame data, determine a category of the first confidence parameter, when the first confidence parameter is fifth preset data, perform an operation on the fifth preset data and the second confidence parameter according to a second preset algorithm to obtain a confidence level of the prediction frame data, and correspondingly save the confidence level of the prediction frame data, the category data, and actual data of the image in the first list to the first list; and when the first confidence parameter is sixth preset data, recording the second confidence parameter as the confidence coefficient of the prediction frame data, and correspondingly storing the confidence coefficient of the prediction frame data, the category data and the actual data of the image into the first list.

Furthermore, when the first calculating subunit is configured to obtain the intersection area and the union area of the prediction frame data and the image according to an eleventh preset algorithm for the prediction frame data and the actual data of the image acquired by the second acquiring unit, the first calculating subunit is configured to calculate the horizontal coordinate, the vertical coordinate, the length and the width of the prediction frame data and the actual horizontal coordinate, the actual vertical coordinate, the actual length and the actual width in the actual data of the image according to the eleventh preset algorithm to obtain the intersection area and the union area of the prediction frame corresponding to the first current prediction frame data and the second current image;

furthermore, when the first calculating subunit is configured to calculate the intersection area and the union area to obtain the second confidence parameter, the first calculating subunit is specifically configured to calculate the intersection area and the union area to obtain the second confidence parameter according to a first preset algorithm;

furthermore, when the first calculating subunit is configured to perform an operation on the intersection area and the union area to obtain the second confidence parameter, the first calculating subunit is specifically configured to perform a ratio operation on the intersection area and the union area to obtain the second confidence parameter.

Optionally, when the determining, saving and clearing unit is configured to correspondingly save the prediction frame data and the actual coordinate data of the image, which correspond to the maximum confidence obtained from the first list, into the second list, the determining, saving and clearing unit is specifically configured to obtain the prediction frame data and the actual coordinate data of the image, which correspond to the maximum confidence from the first list, and correspondingly save the maximum confidence, the prediction frame data, and the actual coordinate data of the image into the second list;

correspondingly, the obtaining module further comprises a first obtaining unit, and the first obtaining unit is used for obtaining a fourth list from the second list;

correspondingly, the first acquiring unit comprises a first classification subunit, a first saving subunit, a first recording subunit, a comparison subunit, a first acquiring subunit and a second saving subunit;

correspondingly, the first classification subunit is configured to classify the prediction frame data in the second list, and sequentially select one of the obtained classes as a first current class;

correspondingly, the first saving subunit is configured to obtain, from the second list, the prediction frame data corresponding to the first current category obtained by the first classification subunit, the actual data of the image, and the confidence level of the prediction frame data, and save the prediction frame data, the actual data of the image, and the confidence level of the prediction frame data to the third list;

correspondingly, the first marking subunit is used for acquiring the maximum confidence from the third list, marking the prediction frame data corresponding to the maximum confidence, and marking the prediction frame data corresponding to the maximum confidence as the first marking prediction frame data;

correspondingly, the first obtaining subunit is further configured to select, from the third list, one of the predictor data other than the first marked predictor data as the second current predictor data;

correspondingly, the second saving subunit is configured to obtain, from the third list, actual data corresponding to the second current prediction box data; obtaining intersection area and union area of the prediction frame corresponding to the first mark prediction frame data and the prediction frame corresponding to the second current prediction frame data according to the prediction frame data corresponding to the first mark prediction frame data and the prediction frame data of the second current prediction frame; calculating the intersection area and the union area to obtain a third confidence parameter;

correspondingly, the comparing subunit is used for comparing the third confidence parameter obtained by the confidence parameter saving subunit with the third comparison data; marking a second current prediction box in a third list when the third confidence parameter is less than the third comparison data;

correspondingly, the first saving subunit is further configured to, when the first obtaining unit obtains all the prediction frame data selected from the third list except the first marked prediction frame data, correspondingly save the prediction frames marked in the third list, and the corresponding confidence, actual data, prediction frame data, and category data in the fourth list, and clear the third list.

Optionally, the fifth executing unit is configured to calculate deviation data of the predicted frame data and the actual coordinate data according to the predicted frame data and the actual coordinate data, and store the deviation data in the fifth list.

Optionally, the module to be identified by processing includes an eleventh acquiring unit, an eleventh executing unit, and an eleventh outputting unit;

correspondingly, the eleventh obtaining unit is used for dividing the picture into grids of the first preset data according to the first preset data; identifying the picture to be identified by using a deep learning network comprising updated first parameter data according to the first preset data, the second preset data and the third preset data to obtain a prediction frame data table; the prediction frame data table comprises second preset data prediction frame data respectively corresponding to all grids; acquiring grids from the prediction frame data table in sequence;

accordingly, an eleventh performing unit, configured to perform an eleventh operation on the mesh acquired by the tenth acquiring unit:

correspondingly, the eleventh execution unit comprises a twelfth acquisition unit and a twelfth execution unit;

correspondingly, the twelfth obtaining unit is configured to obtain the prediction frame data from the second preset data of prediction frame data corresponding to the grid obtained by the eleventh obtaining unit in sequence, and trigger the twelfth executing unit;

accordingly, a twelfth execution unit configured to execute a twelfth operation on the prediction frame data acquired by the twelfth acquisition unit; a twelfth execution unit, configured to specifically acquire the first confidence parameter from the prediction frame data acquired by the twelfth acquisition unit, and when the first confidence parameter is fifth preset data, mark the prediction frame data in the prediction frame data table;

accordingly, the eleven output unit is used for outputting the prediction frame data of all marks in the prediction frame data table.

Further, the twelfth execution unit further includes: a preservation acquisition subunit and a thirteenth execution operation unit,

Further correspondingly, a save acquisition subunit, configured to save the prediction box data to the twelfth list; acquiring the prediction frame data from the twelfth list in sequence to trigger a thirteenth execution unit;

further correspondingly, a thirteenth performing operation unit for performing a thirteenth operation on the prediction box data;

correspondingly, the thirteenth executing operation unit is specifically configured to acquire all category probability data from the storage acquisition subunit prediction frame data, select the largest category probability data from the acquired all category probability data, determine the category data of the prediction frame data according to the largest category probability data, and store the category data and the prediction frame data in correspondence in the twelve lists; classifying all the prediction frame data in the twelve lists according to all the category data in the twelve lists; acquiring any category from the twelve lists, triggering a fourteenth execution operation unit, and emptying the twelfth list when all the categories are acquired from the twelve lists;

further correspondingly, a fourteenth performing operation unit, configured to perform a fourteenth operation on the category acquired by the thirteenth performing operation unit;

correspondingly, the fourteenth executing operation unit is specifically configured to acquire the prediction frame data corresponding to the category from the twelfth list and store the prediction frame data in the thirteenth list; acquiring the largest category probability data of all category probability data corresponding to the category from the prediction frame data in the thirteenth list; marking the prediction box data corresponding to the maximum category probability data in a thirteenth list; according to an eleventh preset algorithm, calculating the prediction frame data corresponding to the maximum category probability data and other prediction frame data in the thirteenth list respectively to obtain sixth confidence parameters corresponding to other prediction frame data in the thirteenth list respectively; performing a fifteenth operation on each sixth confidence parameter; the fifteenth operation is: the apparatus compares the magnitude of the sixth confidence parameter and the sixth comparison data, and marks the prediction box data in the thirteenth list when the sixth confidence parameter is smaller than the thirteenth confidence data; correspondingly marking the prediction frame data of all marks in the thirteenth list in a prediction frame data table; emptying the thirteenth list;

further, when the fourteenth performing operation unit is specifically configured to calculate, according to an eleventh preset algorithm, the prediction frame data corresponding to the maximum category probability data and other prediction frame data in the thirteenth list respectively to obtain sixth confidence parameters corresponding to the other prediction frame data in the thirteenth list, the fourteenth performing operation unit is specifically configured to acquire, from the thirteenth list, the prediction frame data corresponding to the maximum category probability data and mark the prediction frame data as the labeled prediction frame data; when a prediction box data except the marker prediction box data is obtained from the thirteenth list, the device calculates the marker prediction box data and the second current prediction box data according to an eleventh preset algorithm to obtain an intersection area and a union area, and calculates the intersection area and the union area according to the first preset algorithm to obtain a sixth confidence parameter.

The above description is only for the preferred embodiment of the present invention, but the scope of the present invention is not limited thereto, and any changes or substitutions that can be easily conceived by those skilled in the art within the technical scope of the present invention are included in the scope of the present invention. Therefore, the protection scope of the present invention shall be subject to the protection scope of the claims.

Claims

1. A method for carrying on the image recognition based on deep learning network, characterized by that;

2. The method of claim 1, wherein the grid data comprises a grid abscissa, a grid ordinate, a grid length, and a grid width;

the prediction frame data table comprises fourth data of prediction frame data, and the prediction frame data comprises a prediction frame abscissa, a prediction frame ordinate, a prediction frame length, a prediction frame width, a first confidence parameter and third preset data of category probability data;

the fourth data has a binding relationship with the first preset data and the second preset data.

3. The method of claim 2, wherein the binding relationship is: the fourth data is the product of the first preset data and the second preset data;

the first preset data is in a format of data S.

4. The method of claim 1, wherein the third operation comprises the steps of:

step D01: the device acquires category data of the image from actual coordinate data of the image;

step D02: the device judges whether the category data of the image is the same as the category data of the prediction frame data, if so, the actual coordinate data of the image and the prediction frame data are stored in a same-category image table of the prediction frame; otherwise, ending.

5. The method of claim 1, wherein the fourth operation comprises the steps of:

step D11: the device calculates the data of the prediction frame and the actual data of the image according to an eleven preset algorithm to obtain the intersection area and the union area of the data of the prediction frame and the image; calculating the intersection area and the union area to obtain a second confidence parameter;

step D12: the device acquires a first confidence parameter in the prediction frame data, and determines the confidence of the prediction frame data according to the first confidence parameter, a second confidence parameter, fifth preset data and sixth preset data; and correspondingly storing the confidence coefficient of the prediction frame data, the category data of the prediction frame data and the actual data of the image to a first list.

6. The method according to claim 5, wherein the step D12 is specifically: the device acquires a first confidence parameter in the prediction frame data, determines the category of the first confidence parameter, determines the confidence coefficient of the prediction frame data according to fifth preset data and a second confidence parameter when the first confidence parameter is fifth preset data, and correspondingly stores the confidence coefficient of the prediction frame data, the category data and the actual data confidence coefficient of the image to a first list; and when the first confidence parameter is sixth preset data, determining the confidence coefficient of the prediction frame data according to the second confidence parameter, and correspondingly storing the confidence coefficient of the prediction frame data, the category data and the actual data of the image to a first list.

7. The method according to claim 6, wherein the step D12 is specifically: the device acquires a first confidence parameter in the prediction frame data, determines the category of the first confidence parameter, when the first confidence parameter is fifth preset data, calculates the fifth preset data and the second confidence parameter according to a second preset algorithm to obtain the confidence coefficient of the prediction frame data, and correspondingly stores the confidence coefficient of the prediction frame data, the category data and the actual data of the image in a first list to a first list; and when the first confidence parameter is sixth preset data, recording the second confidence parameter as the confidence coefficient of the prediction frame data, and correspondingly storing the confidence coefficient of the prediction frame data, the category data and the actual data of the image into a first list.

8. The method according to claim 5, wherein in step D11, the device obtains an intersection area and a union area of the prediction frame data and the image according to the prediction frame data and the actual data of the image, specifically: and the cross-set area and the union area of the prediction frame corresponding to the first current prediction frame data and the second current image are obtained by calculating the horizontal coordinate, the vertical coordinate, the length and the width of the prediction frame data and the actual horizontal coordinate, the actual vertical coordinate, the actual length and the actual width of the actual data of the image according to an eleventh preset algorithm.

9. The method according to claim 5, wherein in step D11, the operation on the intersection area and the union area to obtain the second confidence parameter includes: the device calculates the intersection area and the union area according to a first preset algorithm to obtain a second confidence parameter.

10. The method according to claim 9, wherein in step D11, the operation on the intersection area and the union area to obtain the second confidence parameter includes: and the device performs ratio operation on the intersection area and the union area according to a first preset algorithm to obtain a second confidence parameter.

11. The method according to claim 1, wherein in step S1-03, the step of obtaining the prediction frame data corresponding to the maximum confidence from the first list and the actual coordinate data of the image is correspondingly saved in a second list, which includes: acquiring the prediction frame data corresponding to the maximum confidence coefficient and the actual coordinate data of the image from the first list, and correspondingly storing the maximum confidence coefficient, the prediction frame data and the actual coordinate data of the image into a second list; the device obtains a fourth list from the second list;

step D23: the apparatus selects one of the predicted frame data other than the first marked predicted frame data from the third list as a second current predicted frame data;

step D26: the apparatus determines whether there is prediction box data that is not regarded as second current prediction box data in the prediction box data other than the marked prediction box data in the third list, and if so, performs step D27; otherwise, correspondingly storing the prediction box marked in the third list and the corresponding confidence coefficient, actual data, prediction box data and category data into a fourth list, emptying the third list, and executing the step D28;

step D27: the apparatus acquires the next prediction box data from the prediction boxes in the third list other than the first marked prediction box data as the second current prediction box data, and returns to step D24.

Step D28: the apparatus determines whether there is a category that is not regarded as the first current category among the categories in the second list, if so, performs step D29; otherwise, emptying the second list, and executing a fifth operation on each prediction box data in the fourth list;

12. The method of claim 1, wherein the fifth operation comprises the steps of: and the device calculates deviation data of the prediction frame data and the actual coordinate data according to the prediction frame data and the actual coordinate data, and stores the deviation data to a fifth list.

13. The method according to claim 1, wherein in step S3, the processing the picture to be recognized by using the deep learning network containing the first parameter data to obtain the prediction frame data table comprises the following steps:

the step T1: when the picture to be identified is acquired, the device divides the picture into grids of first preset data according to the first preset data; identifying the picture to be identified by using a deep learning network comprising updated first parameter data according to the first preset data, the second preset data and the third preset data to obtain a prediction frame data table; the prediction frame data table comprises second preset data prediction frame data respectively corresponding to all grids; sequentially acquiring grids from the prediction frame data table, and performing an eleventh operation on the grids:

the eleventh operation is: acquiring prediction frame data from second preset data prediction frame data corresponding to the grids in sequence, and executing a twelfth operation on the prediction frame data;

the step T2: the apparatus outputs prediction box data for all of the flags in the prediction box data table.

14. The method of claim 13, wherein in the twelfth operation, the prediction box data is marked in a prediction box data table, replaced with: the device saves the prediction box data to a twelfth list; acquiring prediction box data from a twelfth list in sequence, and executing the thirteenth operation on the prediction box data;

when one piece of prediction frame data is acquired from a twelfth list, the device acquires all the category probability data from the prediction frame data, selects the largest category probability data from the acquired all the category probability data, determines the category data of the prediction frame data according to the largest category probability data, and correspondingly stores the category data and the prediction frame data in the twelfth list; classifying all the prediction frame data in the twelve lists according to all the category data in the twelve lists; acquiring any category from the twelve lists, executing a fourteenth operation, and emptying the twelfth list when all the categories are acquired from the twelve lists;

the fourteen operations are: the device acquires the prediction frame data corresponding to the category from the twelfth list and stores the prediction frame data into the thirteenth list; acquiring the maximum category probability data in all category probability data corresponding to the category from the prediction frame data in the thirteenth list; marking the prediction box data corresponding to the maximum category probability data in a thirteenth list; according to an eleventh preset algorithm, calculating the prediction frame data corresponding to the maximum category probability data and other prediction frame data in the thirteenth list respectively to obtain sixth confidence parameters corresponding to other prediction frame data in the thirteenth list respectively; performing a fifteenth operation on each sixth confidence parameter; the fifteenth operation is: the apparatus compares the magnitude of the sixth confidence parameter and the sixth comparison data, and marks the prediction box data in a thirteenth list when the sixth confidence parameter is less than the thirteenth confidence data; correspondingly marking the prediction frame data of all marks in the thirteenth list in a prediction frame data table; the thirteenth list is emptied.

15. The method according to claim 14, wherein the sixth confidence parameters respectively corresponding to the other prediction box data in the thirteenth list are obtained by respectively calculating the prediction box data corresponding to the maximum class probability data and the other prediction box data in the thirteenth list according to an eleventh preset algorithm, which is specifically:

the device acquires prediction frame data corresponding to the maximum category probability data from the thirteenth list and records the prediction frame data as marked prediction frame data; when prediction box data except the marker prediction box data is obtained from the thirteenth list, the device calculates the marker prediction box data and the second current prediction box data according to an eleventh preset algorithm to obtain an intersection area and a union area, and calculates the intersection area and the union area according to the first preset algorithm to obtain a sixth confidence parameter.

16. The device for image recognition based on the deep learning network is characterized by comprising an incoming module, an acquisition module, an execution module, a deviation comparison module, a module to be recognized and an update parameter module;

the execution module comprises a first execution operation unit;

17. The apparatus of claim 16, in which the grid data comprises a grid abscissa, a grid ordinate, a grid length, and a grid width;

18. The apparatus of claim 17, wherein the binding relationship is: the fourth data is the product of the first preset data and the second preset data;

the first preset data is a product of data S and data S.

19. The apparatus of claim 16, wherein the save acquisition unit comprises acquiring as a subunit

The acquiring as subunit is configured to acquire one image from all images corresponding to the grid as an image;

the acquiring subunit is configured to acquire actual coordinate data of the image acquired as the subunit, and acquire category data of the image from the actual coordinate data of the image acquired as the subunit;

the first judging subunit is configured to judge whether the category data of the image obtained as the subunit is the same as the category data of the prediction frame data, and if so, trigger the saving subunit;

the storage subunit is configured to store the actual coordinate data of the image obtained as the subunit and the prediction frame data into a prediction frame same-class image table, and trigger the second judgment subunit, otherwise trigger the second judgment subunit;

the second judging subunit is configured to judge whether a next image can be obtained from all images corresponding to the grid as an image, and if so, obtain the next image from all images corresponding to the grid as an image, and trigger the obtaining subunit; and if not, acquiring images from the same category image tables corresponding to the prediction frames in sequence, and triggering the fourth operation execution unit.

20. The apparatus of claim 16, wherein the fourth unit of performing operations comprises a first computing subunit and a first saving subunit;

the first calculating subunit is configured to obtain an intersection area and a union area of the prediction frame data and the image according to an eleventh preset algorithm for the prediction frame data and the actual data of the image acquired by the second acquiring unit; calculating the intersection area and the union area to obtain a second confidence parameter;

the first storage subunit is configured to obtain a first confidence parameter in the prediction frame data, and determine the confidence of the prediction frame data according to the first confidence parameter, the second confidence parameter of the first calculation subunit, fifth preset data, and sixth preset data; and correspondingly storing the confidence coefficient of the prediction frame data, the category data of the prediction frame data and the actual data of the image to a first list.

21. The apparatus according to claim 20, wherein the first saving subunit is specifically configured to obtain a first confidence parameter in the predicted frame data, determine a category of the first confidence parameter, determine a confidence level of the predicted frame data according to a fifth preset data and a second confidence parameter when the first confidence parameter is the fifth preset data, and correspondingly save the confidence level of the predicted frame data, the category data, and the confidence level of the actual data of the image in a first list; and when the first confidence parameter is sixth preset data, determining the confidence coefficient of the prediction frame data according to the second confidence parameter, and correspondingly storing the confidence coefficient of the prediction frame data, the category data and the actual data of the image to a first list.

22. The apparatus according to claim 21, wherein the first saving subunit is specifically configured to obtain a first confidence parameter in the prediction frame data, determine a category of the first confidence parameter, when the first confidence parameter is fifth preset data, perform an operation on the fifth preset data and the second confidence parameter according to a second preset algorithm to obtain a confidence level of the prediction frame data, and correspondingly save the confidence level of the prediction frame data, the category data, and the actual data of the image in a first list to a first list; and when the first confidence parameter is sixth preset data, recording the second confidence parameter as the confidence coefficient of the prediction frame data, and correspondingly storing the confidence coefficient of the prediction frame data, the category data and the actual data of the image into a first list.

23. The apparatus according to claim 21, wherein when the first calculating subunit is configured to obtain the intersection area and the union area of the predicted box data and the image according to an eleven preset algorithm for the predicted box data and the actual data of the image acquired by the second acquiring unit, the first calculating subunit is configured to calculate the predicted box abscissa, the predicted box ordinate, the predicted box length, and the predicted box width of the predicted box data and the actual abscissa, the actual ordinate, the actual length, and the actual width in the actual data of the image according to the eleven preset algorithm to obtain the intersection area and the union area of the predicted box corresponding to the first current predicted box data and the second current image.

24. The apparatus according to claim 21, wherein when the first calculating subunit is configured to calculate the intersection area and the union area to obtain the second confidence parameter, the first calculating subunit is specifically configured to calculate the intersection area and the union area to obtain the second confidence parameter according to a first preset algorithm.

25. The apparatus according to claim 24, wherein the first computing subunit is further configured to perform a ratio operation on the intersection area and the union area to obtain the second confidence parameter, when the first computing subunit is configured to perform an operation on the intersection area and the union area to obtain the second confidence parameter.

26. The apparatus according to claim 16, wherein when the determining, saving and clearing unit is configured to correspondingly save the prediction frame data and the actual coordinate data of the image, which correspond to the maximum confidence level, obtained from the first list, into the second list, the determining, saving and clearing unit is specifically configured to obtain the prediction frame data and the actual coordinate data of the image, which correspond to the maximum confidence level, from the first list, and correspondingly save the maximum confidence level, the prediction frame data, and the actual coordinate data of the image into the second list;

the acquisition module further comprises a first acquisition unit, wherein the first acquisition unit is used for acquiring a fourth list from the second list;

the first acquisition unit comprises a first classification subunit, a first storage subunit, a first recording subunit, a comparison subunit, a first acquisition subunit and a second storage subunit;

the first classification subunit is configured to classify the prediction frame data in the second list, and select one of the obtained classes in sequence as a first current class;

the first saving subunit is configured to obtain, from the second list, prediction frame data corresponding to the first current category obtained by the first classification subunit, actual data of the image, and a confidence level of the prediction frame data, and save the prediction frame data, the actual data of the image, and the confidence level of the prediction frame data to a third list;

the first marking subunit is used for acquiring the maximum confidence from the third list, marking the prediction frame data corresponding to the maximum confidence, and marking the prediction frame data corresponding to the maximum confidence as the first marking prediction frame data;

the first obtaining subunit is further configured to select, from the third list, one of the prediction frame data other than the first marked prediction frame data as a second current prediction frame data;

the second saving subunit is configured to obtain, from the third list, actual data corresponding to second current prediction frame data; obtaining intersection area and union area of the prediction frame corresponding to the first mark prediction frame data and the prediction frame corresponding to the second current prediction frame data according to the prediction frame data corresponding to the first mark prediction frame data and the prediction frame data of the second current prediction frame; calculating the intersection area and the union area to obtain a third confidence parameter;

the comparing subunit is configured to compare the third confidence parameter obtained by the confidence parameter saving subunit with the third comparison data; marking a second current prediction box in a third list when the third confidence parameter is less than the third comparison data;

the first saving subunit is further configured to, when the first obtaining unit obtains all the prediction frame data selected from the third list except the first marked prediction frame data, correspondingly save the prediction frames marked in the third list and the corresponding confidence degrees, actual data, prediction frame data, and category data in the fourth list, and empty the third list.

27. The apparatus according to claim 16, wherein the fifth execution operation unit is configured to calculate deviation data of the predicted frame data and the actual coordinate data from the predicted frame data and the actual coordinate data, and store the deviation data in a fifth list.

28. The apparatus of claim 16, wherein the processing to-be-recognized module comprises an eleventh obtaining unit, an eleventh executing unit, and the eleventh outputting unit;

the eleventh acquiring unit is configured to divide the picture into first preset data grids according to the first preset data; identifying the picture to be identified by using a deep learning network comprising updated first parameter data according to the first preset data, the second preset data and the third preset data to obtain a prediction frame data table; the prediction frame data table comprises second preset data prediction frame data respectively corresponding to all grids; acquiring grids from the prediction frame data table in sequence;

the eleventh performing unit, configured to perform an eleventh operation on the mesh acquired by the tenth acquiring unit:

the eleventh execution unit comprises a twelfth acquisition unit and a twelfth execution unit;

the twelfth obtaining unit is configured to obtain prediction frame data sequentially from second preset data pieces of prediction frame data corresponding to the grid obtained by the eleventh obtaining unit, and trigger the twelfth executing unit;

the twelfth execution unit is configured to execute a twelfth operation on the prediction frame data acquired by the twelfth acquisition unit; the twelfth execution unit is specifically configured to acquire a first confidence parameter from the prediction frame data acquired by the twelfth acquisition unit, and when the first confidence parameter is fifth preset data, mark the prediction frame data in a prediction frame data table;

and the eleven output unit is used for outputting the prediction frame data of all marks in the prediction frame data table.

29. The apparatus of claim 28, wherein the twelfth execution unit further comprises: a preservation acquisition subunit and a thirteenth execution operation unit,

The storage and acquisition subunit is configured to store the prediction box data in a twelfth list; acquiring prediction frame data from a twelfth list in sequence to trigger the thirteenth execution unit;

the thirteenth performing operation unit is configured to perform the thirteenth operation on the prediction box data;

the thirteenth execution operation unit is specifically configured to acquire all category probability data from the prediction frame data of the storage and acquisition subunit, select the largest category probability data from the acquired all category probability data, determine the category data of the prediction frame data according to the largest category probability data, and store the category data and the prediction frame data in a twelve-row list in a corresponding manner; classifying all the prediction frame data in the twelve lists according to all the category data in the twelve lists; acquiring any category from the twelve lists, triggering the fourteenth execution operation unit, and emptying the twelfth list when all the categories are acquired from the twelve lists;

the fourteenth execution operation unit is configured to execute a fourteenth operation on the category acquired by the thirteenth execution operation unit;

the fourteenth execution operation unit is specifically configured to acquire the prediction box data corresponding to the category from the twelfth list and store the prediction box data in the thirteenth list; acquiring the maximum category probability data in all category probability data corresponding to the category from the prediction frame data in the thirteenth list; marking the prediction box data corresponding to the maximum category probability data in a thirteenth list; according to an eleventh preset algorithm, calculating the prediction frame data corresponding to the maximum category probability data and other prediction frame data in the thirteenth list respectively to obtain sixth confidence parameters corresponding to other prediction frame data in the thirteenth list respectively; performing a fifteenth operation on each sixth confidence parameter; the fifteenth operation is: the apparatus compares the magnitude of the sixth confidence parameter and the sixth comparison data, and marks the prediction box data in a thirteenth list when the sixth confidence parameter is less than the thirteenth confidence data; correspondingly marking the prediction frame data of all marks in the thirteenth list in a prediction frame data table; the thirteenth list is emptied.

30. The apparatus according to claim 29, wherein when the fourteenth performing operation unit is specifically configured to calculate, according to an eleventh preset algorithm, the prediction box data corresponding to the maximum class probability data and other prediction box data in the thirteenth list respectively to obtain sixth confidence parameters corresponding to other prediction box data in the thirteenth list, the fourteenth performing operation unit is specifically configured to acquire the prediction box data corresponding to the maximum class probability data from the thirteenth list and record the prediction box data as labeled prediction box data; when prediction box data except the marker prediction box data is obtained from the thirteenth list, the device calculates the marker prediction box data and the second current prediction box data according to an eleventh preset algorithm to obtain an intersection area and a union area, and calculates the intersection area and the union area according to the first preset algorithm to obtain a sixth confidence parameter.