CN110188720A - A kind of object detection method and system based on convolutional neural networks - Google Patents
A kind of object detection method and system based on convolutional neural networks Download PDFInfo
- Publication number
- CN110188720A CN110188720A CN201910483258.4A CN201910483258A CN110188720A CN 110188720 A CN110188720 A CN 110188720A CN 201910483258 A CN201910483258 A CN 201910483258A CN 110188720 A CN110188720 A CN 110188720A
- Authority
- CN
- China
- Prior art keywords
- neural networks
- convolutional neural
- training
- networks model
- target
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V20/00—Scenes; Scene-specific elements
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- Evolutionary Computation (AREA)
- Life Sciences & Earth Sciences (AREA)
- Artificial Intelligence (AREA)
- General Engineering & Computer Science (AREA)
- Computing Systems (AREA)
- Software Systems (AREA)
- Molecular Biology (AREA)
- Computational Linguistics (AREA)
- Biophysics (AREA)
- Biomedical Technology (AREA)
- Mathematical Physics (AREA)
- General Health & Medical Sciences (AREA)
- Health & Medical Sciences (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Evolutionary Biology (AREA)
- Multimedia (AREA)
- Image Analysis (AREA)
Abstract
The invention discloses object detection method and system based on convolutional neural networks, method includes: to construct the data set for being detected target；According to preset ratio cut partition it is that training set, test set and verifying collect by the image data in the data set of detected target, and image therein is labeled；The network structure of convolutional neural networks model is constructed, the convolutional neural networks model predicts object using different characteristic dimensions；Training set is loaded into convolutional neural networks model to be trained；In the training process, load verifying collection optimizes the parameter of convolutional neural networks model by the method for multiple-authentication；Convolutional neural networks model is tested for the property by test set, detects its generalization ability；Target identification is carried out using the convolutional neural networks model that generalization ability is met the requirements.Train the convolutional neural networks model obtained that can quick and precisely identify the Small object in market, compact intensive or high superposed target through the invention.
Description
Technical field
The present invention relates to field of target recognition more particularly to a kind of object detection method based on convolutional neural networks and it is
System.
Background technique
In recent years with the development of artificial intelligence technology deep learning, after especially proposing convolutional neural networks technology,
The method of target detection has obtained duration update, and detection effect has also obtained qualitative leap.
In view of the rapid development of existing target detection technique, therefore in megastore, there has also been be widely applied.Example
Such as, it needs to carry out unmanned retail identification merchandise classification in market, carries out recognition of face matching, etc. in inlet.Therefore big
In type market, target detection technique has extraordinary development prospect.It in existing technology, can be fast after giving input picture
The jobbie in an image or multiple large-sized objects are detected fastly and mark out position and the type of object.It can be fast
The object detection and recognition end to end of realization one of speed.
In existing target detection technique, detection speed can substantially meet demand, but accuracy rate and identification
There is also biggish defects for effect.The existing target detection technique of detection effect in to(for) single body is preferable, but if needing
When detecting the multiple targets of compact intensive perhaps high superposed target or object to be detected shared overall ratio in figure is smaller
When can not quickly and accurately reach desired target detection effect.This defect is especially tight in the scene application of megastore
Weight, commodity are not of uniform size, and put intensively, and mobile population is huge, and existing image detecting technique is unable to satisfy in this scene
In application.And existing target detection technique generalization ability is weaker, cannot reach desired effect to object classification.If
When the new uncommon length-width ratio and other situations that same type objects occur, it is also unable to complete ideal detection effect.
Summary of the invention
In order to solve the above technical problems, the present invention provides a kind of object detection method based on convolutional neural networks and is
System, specifically, technical scheme is as follows:
On the one hand, the invention discloses a kind of object detection methods based on convolutional neural networks, comprising: building is detected
The data set of target, comprising being detected the image data of target in the data set of the detected target；By the detected mesh
Image data in target data set to the training set of the detected target, test set and is tested according to preset ratio cut partition
Card collection；Target object in all image datas of the training set, test set and verifying collection is labeled；Construct convolution mind
Network structure through network model；The convolutional neural networks model predicts object using different characteristic dimensions；Described
The training set is loaded into convolutional neural networks model to be trained；Process is trained in the convolutional neural networks model
In, the verifying collection is loaded, by the method for multiple-authentication, optimizes the parameter of the convolutional neural networks model；By described
Test set is tested for the property the convolutional neural networks model, detects the generalization ability of the convolutional neural networks model；
Target identification is carried out using the convolutional neural networks model that generalization ability is met the requirements.
It is preferably, described that be loaded into the training set in the convolutional neural networks model and be trained include: setting instruction
Practice parameter；The training parameter includes initial learning rate, weight attenuation rate, maximum number of iterations；By the figure in the training set
Loop iteration training is carried out as data are input in the convolutional neural networks model；After training, training result is extracted.
Preferably, in the training convolutional neural networks model, the loss function calculation formula of use is as follows:
Lossfunction=bool* (2-areaPred) * bce+bool* (2-areaPred)
*(whtrue-whpred)2+bool*bce+(1-bool)*bce*ignore
+bool*bce
Wherein: bool is confidence level；Bce is two-value cross entropy；AreaPred is prediction block range；Whtrue is detected
The length and width true value of target in the picture；Whpred is to be detected the length and width predicted value of target in the picture；Ignore is to hand over simultaneously
Ignore point lower than the object of certain threshold value than (IOU).
Preferably, described during the convolutional neural networks model is trained, the verifying collection is loaded, by more
The method of re-examination card, the parameter for optimizing the convolutional neural networks model includes: in the convolutional neural networks model training mistake
Cheng Zhong loads the verifying collection；Using the method for multiple-authentication, to the loss during the convolutional neural networks model training
Functional value is recorded；According to the loss function value, the connection weight in the neural network model between each neuron is adjusted
Weight；Whether the current training of judgement meets preset trained termination condition；When current training is unsatisfactory for preset trained termination condition
When, continue to be trained the neural network model；When current training meets preset trained termination condition, stop to institute
The training of convolutional neural networks model is stated, each parameter of the convolutional neural networks model after saving training.
Preferably, the preset trained termination condition are as follows: the loss function value reaches preset loss function threshold value
Or frequency of training reaches preset maximum number of iterations.
Preferably, the 0-95 layer of the convolutional neural networks is depth convolutional layer, is made of convolution block and single layer convolutional layer；
The convolution block includes two convolutional layers and a residual error layer；A single layer convolutional layer is inserted into after per several convolution blocks；The volume
The 96-126 layer of product neural network is characterized alternation of bed, includes three scales, is realized by way of convolution kernel in each scale
The feature interaction of part.
On the other hand, the invention also discloses a kind of object detection systems based on convolutional neural networks, comprising: data set
Module is constructed, for constructing the data set of detected target, includes detected target in the data set of the detected target
Image data；Data set division module, for by the image data in the data set of the detected target according to preset ratio
Example is divided to the training set, test set and verifying collection of the detected target；Labeling module, for the training set, test
Target object in all image datas of collection and verifying collection is labeled；Model construction module, for constructing convolutional Neural net
The network structure of network model；The convolutional neural networks model predicts object using different characteristic dimensions；Training module is used
It is trained in being loaded into the training set in the convolutional neural networks model；Optimization module is verified, in the convolution
During neural network model is trained, the verifying collection is loaded, by the method for multiple-authentication, optimizes the convolutional Neural
The parameter of network model；Test module, for being tested for the property by the test set to the convolutional neural networks model,
Detect the generalization ability of the convolutional neural networks model；Target identification module, the volume for being met the requirements by generalization ability
Neural network model is accumulated to carry out target identification.
Preferably, the training module includes: parameter setting submodule, for training parameter to be arranged；The training parameter
Including initial learning rate, weight attenuation rate, maximum number of iterations；Training submodule, for by the picture number in the training set
Loop iteration training is carried out according to being input in the convolutional neural networks model；As a result extraction module mentions after training
Take training result.
Preferably, the verifying optimization module includes: load submodule, in the convolutional neural networks model training
In the process, the verifying collection is loaded；Verify submodule, the method for using multiple-authentication, to the convolutional neural networks mould
Loss function value in type training process is recorded；Weight adjusting submodule, for adjusting institute according to the loss function value
State the connection weight in neural network model between each neuron；Judging submodule, for judging currently whether training meets pre-
If training termination condition；When current training is unsatisfactory for preset trained termination condition, the training module continues to described
Neural network model is trained；Parameter sub-module stored, for when current training meets preset trained termination condition, institute
Training of the training module stopping to the convolutional neural networks model is stated, the convolutional neural networks model after saving training
Each parameter.
Preferably, the 0-95 layer of the convolutional neural networks is depth convolutional layer, is made of convolution block and single layer convolutional layer；
The convolution block includes two convolutional layers and a residual error layer；A single layer convolutional layer is inserted into after per several convolution blocks；The volume
The 96-126 layer of product neural network is characterized alternation of bed, includes three scales, is realized by way of convolution kernel in each scale
The feature interaction of part.
Present invention improves over whole convolutional neural networks structures, carry out operation with multilayer convolutional layer, and increase net
The feature interaction number of plies of network is realized using multiple dimensioned convolution algorithm, improves image characteristics extraction precision, and operation essence is improved
Degree is improved largely to Small object, compact intensive or the multiple targets of high superposed target detection effect.
Detailed description of the invention
To describe the technical solutions in the embodiments of the present invention more clearly, make required in being described below to embodiment
Attached drawing is briefly introduced, it should be apparent that, drawings in the following description are only some embodiments of the invention, for this
For the those of ordinary skill in field, without any creative labor, it can also be obtained according to these attached drawings
His attached drawing.
Fig. 1 is that the present invention is based on the flow charts of the object detection method embodiment of convolutional neural networks；
Fig. 2 is that the present invention is based on the flow charts of another embodiment of the object detection method of convolutional neural networks；
Fig. 3 is the network architecture schematic diagram of convolutional neural networks model in the present invention；
Fig. 4 is convolutional layer schematic diagram of internal structure；
Fig. 5 is residual error layer schematic diagram of internal structure；
Fig. 6 is that the present invention is based on the structural block diagrams of the object detection system embodiment of convolutional neural networks；
Fig. 7 is that the present invention is based on the structural block diagrams of another embodiment of the object detection system of convolutional neural networks.
Appended drawing reference:
A-- depth convolutional layer；B-- feature interaction layer；100-- data set constructs module；200-- data set division module；
300-- labeling module；400-- model construction module；500-- training module；600-- verifies optimization module；700-- tests mould
Block；800-- target identification module；510-- parameter setting submodule；520-- trains submodule；530-- result extracting sub-module；
610-- loads submodule；620-- verifies submodule；630-- weight adjusting submodule；640-- judging submodule；650-- parameter
Sub-module stored.
Specific embodiment
In being described below, for illustration and not for limitation, the tool of such as particular system structure, technology etc is proposed
Body details, so as to provide a thorough understanding of the present application embodiment.However, it will be clear to one skilled in the art that there is no these specific
The application also may be implemented in the other embodiments of details.In other cases, it omits to well-known system, device, electricity
The detailed description of road and method, so as not to obscure the description of the present application with unnecessary details.
It should be appreciated that ought use in this specification and in the appended claims, term " includes " indicates the description
Feature, entirety, step, operation, the presence of element and/or component, but one or more other features, entirety, step are not precluded
Suddenly, the presence or addition of operation, element, component and/or set.
To make simplified form, part related to the present invention is only schematically shown in each figure, they are not represented
Its practical structures as product.In addition, there is identical structure or function in some figures so that simplified form is easy to understand
Component only symbolically depicts one of those, or has only marked one of those.Herein, "one" is not only indicated
" only this ", can also indicate the situation of " more than one ".
It will be further appreciated that the term "and/or" used in present specification and the appended claims is
Refer to any combination and all possible combinations of one or more of associated item listed, and including these combinations.
In order to more clearly explain the embodiment of the invention or the technical proposal in the existing technology, Detailed description of the invention will be compareed below
A specific embodiment of the invention.It should be evident that drawings in the following description are only some embodiments of the invention, for
For those of ordinary skill in the art, without creative efforts, it can also be obtained according to these attached drawings other
Attached drawing, and obtain other embodiments.
The invention discloses a kind of object detection method based on convolutional neural networks, embodiment 1 is as shown in Figure 1, comprising:
S101 constructs the data set for being detected target, includes detected target in the data set of the detected target
Image data；
Training convolutional neural networks model needs many training samples, therefore, it is necessary first to which acquisition contains detected target
Image data, for example if it is intended to the convolutional neural networks model can recognize pen, then just needing to construct a pen
The picture of a large amount of various pens is contained in data set, the inside.Convolutional neural networks model subsequently through a large amount of pen figure
The study of piece can extract the feature of pen, from these pictures so as to identify the pen in various scenes in the later period is applied.
Create the data set for being detected target.For example the data set that the present embodiment uses is the root from ImagNet database
According to the image that the target category of required detection is chosen, such as: teacup, chair, people etc..(ImagNet is one for vision pair
It is the maximum database of image recognition in the world as the large-scale visible database of identification software research).
Image data in the data set of the detected target is detected according to preset ratio cut partition to described by S102
Survey the training set, test set and verifying collection of target；
Specifically, the ratio of division can root building data set and then dividing the picture in the data set
According to being preset, general training set accounts about 80%, the image data of test set and each station 10% or so of verifying collection.Compared with
Good, in the present embodiment, 78% image data in the data set of detected target can be divided in training set；By 10%
Image data be divided in test set, by 12% image data be divided to verifying concentrate.
S103 is labeled the target object in all image datas of the training set, test set and verifying collection；
Specifically, to allow convolutional neural networks that target object can be recognized accurately, then we must first tell him
Which is target object in picture, and especially using the recognition detection with the target object in each market, what is acquired is various
Various articles are may included in the picture of market, then the target item for needing to detect in the picture is what name
Word requires to mark out and in which of picture region etc..That is, to the target in image data in this step
Object is labeled including marking target object title and location region in the image data.
In general, being needed soft using marking to needing detected target to be labeled work in all image datas
Part.Main purpose is to collect the location coordinate information (x, y, w, h) for being detected target in image.(x, y) is to be detected mesh
Position coordinates of the target central point in figure, (w, h) are the length being detected target in the picture and width.
S104 constructs the network structure of convolutional neural networks model；The convolutional neural networks model uses different spies
Scale is levied to predict object；
Specifically, in order to identify various Small object objects, dense object, the even object of lap
Body then just has higher requirement to the network architecture of convolutional neural networks model.
The present embodiment is using the algorithm of target detection end to end based on deep learning.Specifically, by target area
Prediction and target category prediction are integrated in single Neural model, realize the real-time quick mesh in the higher situation of accuracy rate
Mark detection and identification.It can be by whole figure as the defeated of network using the convolutional neural networks model (after training) in the present embodiment
Enter, the position of boundingbox and the classification of boundingbox are directly returned in output layer.
The convolutional neural networks used in the present embodiment predict object using different characteristic dimensions, increase network
The feature interaction number of plies, by improving whole convolutional neural networks structure on the basis of traditional convolutional neural networks, with more
Layer convolutional layer carries out operation, is realized using multiple dimensioned convolution algorithm, improves image characteristics extraction precision, improves operation essence
Degree is improved largely to Small object, compact intensive or the multiple targets of high superposed target detection effect.
S105 is loaded into the training set in the convolutional neural networks model and is trained；
After the framework for building convolutional neural networks model, then training set is loaded into the model and is trained.
S106 loads the verifying collection, passes through multiple-authentication during the convolutional neural networks model is trained
Method, optimize the parameter of the convolutional neural networks model；
Specifically, in the training process, it is also necessary to carry out the parameter to trained convolutional neural networks model in conjunction with verifying collection
It is adjusted.Using the method for multiple-authentication, verifying collection is loaded into test
Card, the loss function value for calculating at this stage be it is how many, adjust convolutional Neural net in turn further according to calculating acquisition loss function value
The parameter of network.It then proceedes to train, then verifies, then adjust, retraining, loop back and forth like this, by way of multiple-authentication, one
Optimize the parameter of convolutional neural networks model step by step.
S107 is tested for the property the convolutional neural networks model by the test set, detects the convolution mind
Generalization ability through network model；
After training, then detect by the image data in test set the property of the convolutional neural networks model after training
Can, detect its generalization ability.Generalization ability (generalization ability) refers to machine learning algorithm to fresh sample
Adaptability.The destination of study is to acquire the rule for lying in data behind, other than the study collection with same rule
Data, trained network can also provide suitable output, which is known as generalization ability.
S108 carries out target identification using the convolutional neural networks model that generalization ability is met the requirements.
Specifically, can be protected if detecting that the generalization ability of the convolutional neural networks model is met the requirements
Each parameter for depositing the convolutional neural networks model carries out the knowledge such as Small object, intensive target by the convolutional neural networks model
Not.
In the present embodiment, the data set for being ready to complete to classify and mark is (comprising training set, verifying collection and test
Collection), training set is loaded into and is trained in the convolutional neural networks model built, load verifying collection carries out in training process
Verifying.Training result data are loaded after completing training.Test set picture is imported in test program to be tested, and test knot is obtained
Fruit.The present invention has preferable resolution for Small object and compact target.
In above-described embodiment, step S105 is loaded into the training set in the convolutional neural networks model and is trained packet
It includes:
Training parameter is arranged in S1051；The training parameter includes initial learning rate, weight attenuation rate, greatest iteration time
Number；
Specifically, needing first to carry out initial setting up to model before training starts, in general, needing first to be arranged initial
Learning efficiency, for example, initial learning rate 0.1 is arranged, 4 times of decline every time terminates when to 0.0005；In addition, there is provision of weight
A series of training parameters such as attenuation rate and maximum number of iterations.
Image data in the training set is input in the convolutional neural networks model and is recycled repeatedly by S1052
Generation training；
S1053 after training, extracts training result.
Specifically, after training, the parameters such as weight of convolutional neural networks model after extracting training, after being trained
Convolutional neural networks model.
Preferably, loss function equally plays very important effect in the training convolutional neural networks model.
It is main foundation in learning process is also to judge that the important of algorithm superiority and inferiority is sentenced after study by calculating the size of loss function
According to.Common loss function has mean square error (MSE), and it is also simplest loss function which, which is most classical, almost
It is omnipotent, but it is less accurate.The loss function that the present invention uses is by being detected the loss of the centre coordinate of target, being detected target
The loss of length and width regressand value, the loss of preceding background confidence level and classification in the picture is lost this four losses and is constituted.
Specifically, being detected the location coordinate information (x, y, w, h) of target in such as certain image.Wherein, (x, y) is tested
Position coordinates of the central point of target in figure are surveyed, (w, h) is the length being detected target in the picture and width.Each costing bio disturbance is such as
Under:
(1) loss of xy (object centre coordinate): xyloss=bool* (2-areaPred) * bce, wherein bool is to set
Reliability, wherein bce is that the two-value of xy value intersects entropy loss, and it is smaller that this is worth smaller entire penalty values.
(2) loss of wh (anchor length and width regressand value) is calculated:
Whloss=bool* (2-areaPred) * (whtrue-whpred)2, the case where ensuring confidence level (bool)
Under, areaPred needs are bigger, and wh is needed as close to true value wh.
(3) confidence level loss (preceding background) loss: bool*bce+ (1-bool) * bce*ignore is calculated, wherein bool is
Confidence level, bce are the two-value cross entropy of predicted value and actual degree of belief, and ignore indicates iou lower than certain threshold value but certain
Existing object, ignores.In the case where ensuring confidence level (bool), predicted value is needed as close to true value, together
When there is no the part of object to need as close to background true value, while multiplied by needing to ignore point accordingly.
(4) classification loss: bool*bce is calculated, confidence level is multiplied by a polytypic cross entropy.
To sum up, it sums to above four losses, the loss function of the invention used can be obtained, specific calculation formula is such as
Under:
Lossfunction=bool* (2-areaPred) * bce+bool* (2-areaPred)
*(whtrue-whpred)2+bool*bce+(1-bool)*bce*ignore
+bool*bce
Wherein: bool is confidence level；Bce is two-value cross entropy；AreaPred is prediction block range；Whtrue is detected
The length and width true value of target in the picture；Whpred is to be detected the length and width predicted value of target in the picture；Ignore is to hand over simultaneously
Ignore point than the object lower than certain threshold value.
Another embodiment of the method for the present invention, as shown in Figure 2, comprising:
S201 constructs the data set for being detected target, includes detected target in the data set of the detected target
Image data；
Image data in the data set of the detected target is detected according to preset ratio cut partition to described by S202
Survey the training set, test set and verifying collection of target；
S203 is labeled the target object in all image datas of the training set, test set and verifying collection；
S204 constructs the network structure of convolutional neural networks model；The convolutional neural networks model uses different spies
Scale is levied to predict object；
S205 is loaded into the training set in the convolutional neural networks model and is trained；
Specifically, initial learning rate, weight attenuation rate, a system such as maximum number of iterations is first arranged before training starts
Column training parameter.Then start to train again.
S206 loads the verifying collection during convolutional neural networks model training；
Therefore the verifying for being also trained the stage in training process by verifying collection waits training to terminate, extract training
As a result (each parameter, such as each neuron weight, learning rate etc. of convolutional neural networks model at the end of training).
S207, using the method for multiple-authentication, to the loss function value during the convolutional neural networks model training
It is recorded；
S208 adjusts the connection weight in the neural network model between each neuron according to the loss function value；
S209, judges whether current training meets preset trained termination condition；If so, entering step S210；Otherwise, it returns
Step S205 is returned to continue to be trained the neural network model；
Specifically, preset trained termination condition can sets itself, for example can be loss function value and reach preset damage
Lose function threshold；Or frequency of training reaches preset maximum number of iterations etc..
S210 stops the training to the convolutional neural networks model, the convolutional neural networks mould after saving training
Each parameter of type；
S211 is tested for the property the convolutional neural networks model by the test set, detects the convolution mind
Generalization ability through network model；
It is tested using test set: by the training result after training (for example, convolutional neural networks model after training
The weight of current each neuron) it is loaded into training pattern, load test collection data are tested for the property simultaneously in test code
The generalization ability of detection model, obtains test result.
S212 carries out target identification using the convolutional neural networks model that generalization ability is met the requirements.
In the present embodiment, when training convolutional neural networks model, current training can also be verified,
Specifically, calculating and saving the loss function value of current training pattern, then adjusted in turn according to current loss function value
The weight of each neuron of training pattern, to play the role of optimizing the parameter of the training pattern, so that the training pattern is defeated
Result out is more and more correct, and error is smaller and smaller.When verifying optimization module detects the loss function value of current training pattern
When having reached preset loss function threshold value, then can first deconditioning, save current training result (each parameter), so
The training pattern is tested by test module again afterwards, detects its generalization ability, is trained in this way
Convolutional neural networks model has preferable resolution for Small object and compact target.
In any of the above-described embodiment, the network architecture of the convolutional neural networks model is specific as shown in figure 3, the convolution is refreshing
0-95 layer through network model is depth convolutional layer (part A in figure)；Specifically, the depth convolutional layer of convolutional neural networks model
It is made of convolution residual block (2*conv+1*res) and single layer convolutional layer (conv)；The convolution residual block includes two convolutional layers
With a residual error layer；A single layer convolutional layer is inserted into after per several convolution residual blocks.Specific structure in convolutional layer and residual error layer
As shown in Figure 4, Figure 5.Preferably, in the present embodiment, depth convolutional layer 68 convolutional layers in total, remaining is residual error layer
(residual)。
Can there are one BN layers (BatchNormalization) (carrying out batch standardization) and one after each convolutional layer
ReLU (activation primitive).Specifically, each convolutional layer has 32 convolution kernels (filter) in the present embodiment, each convolution kernel is big
Small is 3*3, step-length 1.The picture for inputting pixel 416*416*3, output obtains 416*416*32's after convolution algorithm
feature map.Residual layers do not influence input and output as a result, i.e. input and output are generally consistent, and main function is to subtract
Small loss controls the propagation of gradient.
78-109 layers be network feature interaction layer (part B in Fig. 3).Three branch's outputs can be obtained by upper layer operation to make
Prediction, the size for exporting characteristic pattern is respectively 13*13, tri- scales of 26*26,52*52, in each scale, passes through convolution kernel
Mode realizes local feature interaction, realizes that the part between feature map is special by the convolution algorithm mode of 3*3 and 1*1
Sign interaction.Wherein: the characteristic pattern of 13*13 uses (116 × 90)；(156×198)；(373 × 326) this 3 anchor；26*26
Characteristic pattern use (30 × 61)；(62×45)；(59 × 119) this 3 anchor；The characteristic pattern of 52*52 uses (10 × 13)；
(16×30)；(33 × 23) this 3 anchor.
Present invention improves over whole convolutional neural networks structures, carry out operation with multilayer convolutional layer, and increase net
The feature interaction number of plies of network is realized using multiple dimensioned convolution algorithm, improves image characteristics extraction precision, and operation essence is improved
Degree is improved largely to Small object, compact intensive or the multiple targets of high superposed target detection effect.
On the other hand, the invention also discloses a kind of object detection system based on convolutional neural networks, the target detections
System can be used it is of the invention target object is detected based on the object detection method of convolutional neural networks, specifically, this implementation
The object detection system based on convolutional neural networks, as shown in Figure 6, comprising:
Data set constructs module 100, for constructing the data set of detected target, in the data set of the detected target
Image data comprising being detected target；Image data set needed for creation.Specifically, the data set that such as the present embodiment uses
It is the image chosen from ImagNet database according to the target category of required detection, such as: teacup, chair, people etc.
(ImagNet is the large-scale visible database for being used for the research of visual object identification software, is that image recognition is maximum in the world
Database).
Data set division module 200, for by the image data in the data set of the detected target according to preset
Ratio cut partition to the training set of the detected target, test set and verifying collects；Specifically, data set division module 200 to from
The image data downloaded in ImageNet database is classified, and is divided into three classes and is distributed as training set, test set, verifying collection.It draws
Point ratio can be with preset in advance, and general training collection proportion is larger, about can achieve 80%, test set and verifying collection ratio
Close, probably each station 10% or so certainly, on this basis, can carry out subtle adjustment again, for example, in the present embodiment, training
Collection, test set, verifying collect the ratio of shared conceptual data collection respectively (78%, 10%, 12%).
Labeling module 300, for the target object in all image datas to the training set, test set and verifying collection
It is labeled；Labeling module 300 is needed to needing detected target to be labeled work in all image datas using mark
Software.Main purpose is to collect the location coordinate information (x, y, w, h) for being detected target in image.(x, y) is detected
Position coordinates of the central point of target in figure, (w, h) are the length being detected target in the picture and width.
Model construction module 400, for constructing the network structure of convolutional neural networks model；The convolutional neural networks mould
Type predicts object using different characteristic dimensions；Specifically, in order to identify various Small object objects, dense object,
The even target object of lap then just has higher requirement to the network architecture of convolutional neural networks model.This implementation
Example is improved on the basis of traditional convolutional neural networks, allows the convolutional neural networks model using difference
Characteristic dimension predict object.
The present embodiment improves whole convolutional neural networks structure, carries out operation with multilayer convolutional layer, and increase
The feature interaction number of plies of network is realized using multiple dimensioned convolution algorithm, improves image characteristics extraction precision, and operation essence is improved
Degree is improved largely to Small object, compact intensive or the multiple targets of high superposed target detection effect
Training module 500 is trained for being loaded into the training set in the convolutional neural networks model；
Optimization module 600 is verified, for loading the verifying during convolutional neural networks model is trained
Collection, by the method for multiple-authentication, optimizes the parameter of the convolutional neural networks model；
Specifically, in the training process, it is also necessary to carry out the parameter to trained convolutional neural networks model in conjunction with verifying collection
It is adjusted.Using the method for multiple-authentication, verifying collection is loaded into test
Card, the loss function value for calculating at this stage be it is how many, adjust convolutional Neural net in turn further according to calculating acquisition loss function value
The parameter of network.It then proceedes to train, then verifies, then adjust, retraining, loop back and forth like this, by way of multiple-authentication, one
Optimize the parameter of convolutional neural networks model step by step.
Test module 700 is detected for being tested for the property by the test set to the convolutional neural networks model
The generalization ability of the convolutional neural networks model；After training, then training detected by the image data in test set
The performance of convolutional neural networks model afterwards, detects its generalization ability.
Target identification module 800, convolutional neural networks model for being met the requirements by generalization ability carry out target
Identification.If detecting that the generalization ability of the convolutional neural networks model is met the requirements, convolution mind can be saved
Each parameter through network model, later period can load each parameter of the convolutional neural networks model of preservation to carry out Small object, intensive
The identification such as target.
Preferably, on the basis of the above system embodiment, as shown in fig. 7, the training module 500 includes:
Parameter setting submodule 510, for training parameter to be arranged；The training parameter includes initial learning rate, and weight declines
Lapse rate, maximum number of iterations；Specifically, needing first to carry out initial setting up to model, in general, needing before training starts
Initial learning efficiency is first set, for example, initial learning rate 0.1 is arranged；In addition, there is provision of weight attenuation rate and greatest iteration time
A series of training parameters such as number.
Training submodule 520, for the image data in the training set to be input to the convolutional neural networks model
Middle progress loop iteration training；
As a result extraction module 530, after training, extraction training result.Specifically, extracting and (protecting after training
Deposit) parameters such as the weight of convolutional neural networks model after training, the convolutional neural networks model after being trained.
Preferably, loss function equally plays very important effect in the training convolutional neural networks model.
It is main foundation in learning process is also to judge that the important of algorithm superiority and inferiority is sentenced after study by calculating the size of loss function
According to.Common loss function has mean square error (MSE), and it is also simplest loss function which, which is most classical, almost
It is omnipotent, but it is less accurate.The loss function that the present invention uses is by being detected the loss of the centre coordinate of target, being detected target
The loss of length and width regressand value, the loss of preceding background confidence level and classification in the picture is lost this four losses and is constituted.
Specifically, being detected the location coordinate information (x, y, w, h) of target in such as certain image.Wherein, (x, y) is tested
Position coordinates of the central point of target in figure are surveyed, (w, h) is the length being detected target in the picture and width.Each costing bio disturbance is such as
Under:
(1) loss of xy (object centre coordinate): xyloss=bool* (2-areaPred) * bce, wherein bool is to set
Reliability, wherein bce is that the two-value of xy value intersects entropy loss, and it is smaller that this is worth smaller entire penalty values.
(2) loss of wh (anchor length and width regressand value) is calculated:
Whloss=bool* (2-areaPred) * (whtrue-whpred)2, the case where ensuring confidence level (bool)
Under, areaPred needs are bigger, and wh is needed as close to true value wh.
(3) confidence level loss (preceding background) loss: bool*bce+ (1-bool) * bce*ignore is calculated, wherein bool is
Confidence level, bce are the two-value cross entropy of predicted value and actual degree of belief, and ignore indicates iou lower than certain threshold value but certain
Existing object, ignores.In the case where ensuring confidence level (bool), predicted value is needed as close to true value, together
When there is no the part of object to need as close to background true value, while multiplied by needing to ignore point accordingly.
(4) classification loss: bool*bce is calculated, confidence level is multiplied by a polytypic cross entropy.
To sum up, it sums to above four losses, the loss function of the invention used can be obtained, specific calculation formula is such as
Under:
Lossfunction=bool* (2-areaPred) * bce+bool* (2-areaPred)
*(whtrue-whpred)2+bool*bce+(1-bool)*bce*ignore
+bool*bce
Wherein: bool is confidence level；Bce is two-value cross entropy；AreaPred is prediction block range；Whtrue is detected
The length and width true value of target in the picture；Whpred is to be detected the length and width predicted value of target in the picture；Ignore is to hand over simultaneously
Ignore point than the object lower than certain threshold value.
Using loss function of the invention, it is more conducive to the training to Small object object identification model.
In above-described embodiment, the verifying optimization module 600 includes:
Submodule 610 is loaded, for during the convolutional neural networks model training, loading the verifying collection；
Verify submodule 620, the method for using multiple-authentication, during the convolutional neural networks model training
Loss function value recorded；
Weight adjusting submodule 630, for adjusting each nerve in the neural network model according to the loss function value
Connection weight between member；
Judging submodule 640, for judging currently whether training meets preset trained termination condition；It is trained not when current
When meeting preset trained termination condition, the training module 500 continues to be trained the neural network model；For example,
Preset trained termination condition are as follows: the loss function value reaches preset loss function threshold value or frequency of training reaches preset
Maximum number of iterations.As long as the loss function value that so verifying submodule 620 authenticates to current convolutional neural networks model reaches
Preset maximum number of iterations will terminate to train before preset loss function value or the number of training have reached.
Parameter sub-module stored 650, for when current training meets preset trained termination condition, the training module
500 stop the training to the convolutional neural networks model, each parameter of the convolutional neural networks model after saving training.
After training, parameter sub-module stored 650 will save each parameter of "current" model, and later period test module 700 passes through test
Previously stored training result (each parameter of the training pattern of preservation) is loaded into training pattern when collection test, in test generation
Load test collection data are tested for the property the generalization ability of simultaneously detection model in code, if detecting the generalization ability of the model
It meets the requirements, then can be carried out in target identification, especially market using the trained convolutional neural networks model
The identification of all kinds of Small objects, intensive target etc..If generalization ability is unsatisfactory for requiring, will return again to convolutional Neural net
Network model is trained, and until its general Huaneng Group power is met the requirements, finally obtains the volume met the requirements by adequately training
Product neural network model, saves its final training result (each parameter of the trained convolutional neural networks model).
The network architecture of the convolutional neural networks model used in the present invention is as shown in figure 3, the convolutional neural networks
0-95 layers are depth convolutional layer (in figure shown in A), and the depth convolutional layer of convolutional neural networks model is by convolution block and single layer convolution
Layer is constituted；The convolution block includes two convolutional layers and a residual error layer；A single layer convolutional layer is inserted into after per several convolution blocks.
Depth convolutional layer in the present embodiment includes 68 convolutional layers, remaining is residual error layer.
Can there are one BN layers (BatchNormalization) (carrying out batch standardization) and one after each convolutional layer
ReLU (activation primitive).Each convolutional layer has 32 convolution kernels (filter), and each convolution kernel size is 3*3, step-length 1.It is defeated
The picture for entering pixel 416*416*3, output obtains the feature map of 416*416*32 after convolution algorithm.Residual
Layer does not influence input and output as a result, i.e. input and output are generally consistent, and main function is to reduce loss, controls the biography of gradient
It broadcasts.
The 96-126 layer of the convolutional neural networks is characterized alternation of bed (in Fig. 3 shown in B), includes three scales, each
Local feature interaction is realized by way of convolution kernel in scale.Three branch's outputs can be obtained by upper layer operation to make prediction,
The size of output characteristic pattern is respectively 13*13, tri- scales of 26*26,52*52, real by way of convolution kernel in each scale
The feature interaction in current situation portion realizes the interaction of the local feature between feature map by the convolution algorithm mode of 3*3 and 1*1.
Wherein: the characteristic pattern of 13*13 uses (116 × 90)；(156×198)；(373 × 326) this 3 anchor；The feature of 26*26
Figure uses (30 × 61)；(62×45)；(59 × 119) this 3 anchor；The characteristic pattern of 52*52 uses (10 × 13)；(16×
30)；(33 × 23) this 3 anchor.
The object detection system based on convolutional neural networks of the present embodiment improves whole convolutional neural networks knot
Structure carries out operation with multilayer convolutional layer, and increases the feature interaction number of plies of network, is realized using multiple dimensioned convolution algorithm,
Image characteristics extraction precision is improved, operational precision is improved, to Small object, the compact intensive or multiple mesh of high superposed target
Target detection effect is improved largely.
The device of the invention embodiment is corresponding with embodiment of the method for the invention, of the invention based on convolutional neural networks
The technical detail of object detection method is equally applicable to the object detecting device of the invention based on convolutional neural networks, to reduce
It repeats, repeats no more.
The present invention be referring to according to the method for the embodiment of the present invention, the process of equipment (system) and computer program product
Figure and/or block diagram describe.It should be understood that every one stream in flowchart and/or the block diagram can be realized by computer program instructions
The combination of process and/or box in journey and/or box and flowchart and/or the block diagram.It can provide these computer programs
Instruct the processor of general purpose computer, special purpose computer, Embedded Processor or other programmable data processing devices to produce
A raw machine, so that being generated by the instruction that computer or the processor of other programmable data processing devices execute for real
The device for the function of being specified in present one or more flows of the flowchart and/or one or more blocks of the block diagram.
These computer program instructions, which may also be stored in, is able to guide computer or other programmable data processing devices with spy
Determine in the computer-readable memory that mode works, so that it includes referring to that instruction stored in the computer readable memory, which generates,
Enable the manufacture of device, the command device realize in one box of one or more flows of the flowchart and/or block diagram or
The function of being specified in multiple boxes.
These computer program instructions also can be loaded onto a computer or other programmable data processing device, so that counting
Series of operation steps are executed on calculation machine or other programmable devices to generate computer implemented processing, thus in computer or
The instruction executed on other programmable devices is provided for realizing in one or more flows of the flowchart and/or block diagram one
The step of function of being specified in a box or multiple boxes.
Although preferred embodiments of the present invention have been described, it is created once a person skilled in the art knows basic
Property concept, then additional changes and modifications may be made to these embodiments.So it includes excellent that the following claims are intended to be interpreted as
It selects embodiment and falls into all change and modification of the scope of the invention.
Obviously, various changes and modifications can be made to the invention without departing from essence of the invention by those skilled in the art
Mind and range.In this way, if these modifications and changes of the present invention belongs to the range of the claims in the present invention and its equivalent technologies
Within, then the present invention is also intended to include these modifications and variations.
Claims (10)
1. a kind of object detection method based on convolutional neural networks characterized by comprising
Construct the data set for being detected target, the image data comprising detected target in the data set of the detected target；
By the image data in the data set of the detected target according to preset ratio cut partition to the detected target
Training set, test set and verifying collection；
Target object in all image datas of the training set, test set and verifying collection is labeled；
Construct the network structure of convolutional neural networks model；The convolutional neural networks model is using different characteristic dimensions come pre-
Survey object；
The training set is loaded into the convolutional neural networks model to be trained；
During the convolutional neural networks model is trained, the verifying collection is loaded, it is excellent by the method for multiple-authentication
Change the parameter of the convolutional neural networks model；
The convolutional neural networks model is tested for the property by the test set, detects the convolutional neural networks model
Generalization ability；
Target identification is carried out using the convolutional neural networks model that generalization ability is met the requirements.
2. a kind of object detection method based on convolutional neural networks according to claim 1, which is characterized in that it is described
It is loaded into the training set in the convolutional neural networks model and is trained and include:
Training parameter is set；The training parameter includes initial learning rate, weight attenuation rate, maximum number of iterations；
Image data in the training set is input in the convolutional neural networks model and carries out loop iteration training；
After training, training result is extracted.
3. a kind of object detection method based on convolutional neural networks according to claim 1, which is characterized in that in training
In the convolutional neural networks model, the loss function calculation formula of use is as follows:
Lossfunction=bool* (2-areaPred) * bce+bool* (2-areaPred) * (whtrue-whpred)2+
bool*bce+(1-bool)*bce*ignore+bool*bce
Wherein: bool is confidence level；Bce is two-value cross entropy；AreaPred is prediction block range；Whtrue is to be detected target
Length and width true value in the picture；Whpred is to be detected the length and width predicted value of target in the picture；Ignore is to hand over and than low
Ignore point in the object of certain threshold value.
4. a kind of object detection method based on convolutional neural networks according to claim 1, which is characterized in that it is described
It during the convolutional neural networks model is trained, loads the verifying and collects, by the method for multiple-authentication, described in optimization
The parameter of convolutional neural networks model includes:
During the convolutional neural networks model training, the verifying collection is loaded；
Using the method for multiple-authentication, the loss function value during the convolutional neural networks model training is recorded；
According to the loss function value, the connection weight in the neural network model between each neuron is adjusted；
Whether the current training of judgement meets preset trained termination condition；
When current training is unsatisfactory for preset trained termination condition, continue to be trained the neural network model；
When current training meets preset trained termination condition, stops the training to the convolutional neural networks model, save
Each parameter of the convolutional neural networks model after training.
5. a kind of object detection method based on convolutional neural networks according to claim 4, which is characterized in that described pre-
If training termination condition are as follows:
The loss function value reaches preset loss function threshold value or frequency of training reaches preset maximum number of iterations.
6. a kind of object detection method based on convolutional neural networks according to claim 1-5, feature exist
In,
The 0-95 layer of the convolutional neural networks is depth convolutional layer, is made of convolution block and single layer convolutional layer；The convolution block
Include two convolutional layers and a residual error layer；A single layer convolutional layer is inserted into after per several convolution blocks；
The 96-126 layer of the convolutional neural networks is characterized alternation of bed, includes three scales, passes through convolution kernel in each scale
Mode realize local feature interaction.
7. a kind of object detection system based on convolutional neural networks characterized by comprising
Data set constructs module, includes quilt in the data set of the detected target for constructing the data set of detected target
Detect the image data of target；
Data set division module, for by the image data in the data set of the detected target according to preset ratio cut partition
To the training set of the detected target, test set and verifying collection；
Labeling module is marked for the target object in all image datas to the training set, test set and verifying collection
Note；
Model construction module, for constructing the network structure of convolutional neural networks model；The convolutional neural networks model uses
Different characteristic dimension predicts object；
Training module is trained for being loaded into the training set in the convolutional neural networks model；
Optimization module is verified, for during the convolutional neural networks model is trained, loading the verifying collection, is passed through
The method of multiple-authentication optimizes the parameter of the convolutional neural networks model；
Test module detects the volume for being tested for the property by the test set to the convolutional neural networks model
The generalization ability of product neural network model；
Target identification module, convolutional neural networks model for being met the requirements by generalization ability carry out target identification.
8. a kind of object detection system based on convolutional neural networks according to claim 7, which is characterized in that the instruction
Practicing module includes:
Parameter setting submodule, for training parameter to be arranged；The training parameter includes initial learning rate, weight attenuation rate, most
Big the number of iterations；
Training submodule, is followed for the image data in the training set to be input in the convolutional neural networks model
Ring iterative training；
As a result extraction module, after training, extraction training result.
9. a kind of object detection system based on convolutional neural networks according to claim 7, which is characterized in that described to test
Demonstrate,proving optimization module includes:
Submodule is loaded, for during the convolutional neural networks model training, loading the verifying collection；
Verify submodule, the method for using multiple-authentication, to the loss during the convolutional neural networks model training
Functional value is recorded；
Weight adjusting submodule, for adjusting in the neural network model between each neuron according to the loss function value
Connection weight；
Judging submodule, for judging currently whether training meets preset trained termination condition；When current training be unsatisfactory for it is pre-
If training termination condition when, the training module continues to be trained the neural network model；
Parameter sub-module stored, for when current training meets preset trained termination condition, the training module stops pair
The training of the convolutional neural networks model, each parameter of the convolutional neural networks model after saving training.
10. according to a kind of described in any item object detection systems based on convolutional neural networks of claim 7-9, feature exists
In,
The 0-95 layer of the convolutional neural networks is depth convolutional layer, is made of convolution block and single layer convolutional layer；The convolution block
Include two convolutional layers and a residual error layer；A single layer convolutional layer is inserted into after per several convolution blocks；
The 96-126 layer of the convolutional neural networks is characterized alternation of bed, includes three scales, passes through convolution kernel in each scale
Mode realize local feature interaction.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910483258.4A CN110188720A (en) | 2019-06-05 | 2019-06-05 | A kind of object detection method and system based on convolutional neural networks |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910483258.4A CN110188720A (en) | 2019-06-05 | 2019-06-05 | A kind of object detection method and system based on convolutional neural networks |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN110188720A true CN110188720A (en) | 2019-08-30 |
Family
ID=67720359
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201910483258.4A Pending CN110188720A (en) | 2019-06-05 | 2019-06-05 | A kind of object detection method and system based on convolutional neural networks |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN110188720A (en) |
Cited By (45)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110569841A (en) * | 2019-09-02 | 2019-12-13 | 中南大学 | contact gateway key component target detection method based on convolutional neural network |
| CN110647622A (en) * | 2019-09-29 | 2020-01-03 | 北京金山安全软件有限公司 | Interactive data validity identification method and device |
| CN110929670A (en) * | 2019-12-02 | 2020-03-27 | 合肥城市云数据中心股份有限公司 | Muck truck cleanliness video identification and analysis method based on yolo3 technology |
| CN111008641A (en) * | 2019-10-24 | 2020-04-14 | 云南电网有限责任公司昆明供电局 | Power transmission line tower external force damage detection method based on convolutional neural network |
| CN111079861A (en) * | 2019-12-31 | 2020-04-28 | 国网北京市电力公司 | Power distribution network voltage abnormity diagnosis method based on image rapid processing technology |
| CN111079683A (en) * | 2019-12-24 | 2020-04-28 | 天津大学 | Remote sensing image cloud and snow detection method based on convolutional neural network |
| CN111127495A (en) * | 2019-12-04 | 2020-05-08 | 深圳市图郅创新科技有限公司 | Neural network model building method for image key point detection |
| CN111191570A (en) * | 2019-12-26 | 2020-05-22 | 国网北京市电力公司 | Image recognition method and device |
| CN111222462A (en) * | 2020-01-07 | 2020-06-02 | 河海大学 | Target detection-based intelligent labeling method for apparent feature monitoring data |
| CN111353504A (en) * | 2020-03-02 | 2020-06-30 | 济南大学 | Source and machine identification method based on image block diversity selection and residual prediction module |
| CN111353976A (en) * | 2020-02-25 | 2020-06-30 | 南京大学 | Sand grain target detection method based on convolutional neural network |
| CN111523713A (en) * | 2020-04-14 | 2020-08-11 | 中国地质大学（北京） | Method and device for predicting residual oil saturation distribution in oil field |
| CN111539178A (en) * | 2020-04-26 | 2020-08-14 | 成都市深思创芯科技有限公司 | Chip layout design method and system based on neural network and manufacturing method |
| CN111833395A (en) * | 2020-06-04 | 2020-10-27 | 西安电子科技大学 | Direction-finding system single target positioning method and device based on neural network model |
| CN111898598A (en) * | 2020-07-03 | 2020-11-06 | 贵州大学 | Target detection method based on text in dynamic scene |
| CN112084950A (en) * | 2020-09-10 | 2020-12-15 | 上海庞勃特科技有限公司 | Target detection method and detection device based on sparse convolutional neural network |
| CN112101449A (en) * | 2020-09-11 | 2020-12-18 | 南京邮电大学 | Method for inverting radar data by satellite cloud picture based on semantic loss |
| CN112098997A (en) * | 2020-09-18 | 2020-12-18 | 欧必翼太赫兹科技(北京)有限公司 | Three-dimensional holographic imaging security inspection radar image foreign matter detection method |
| CN112163541A (en) * | 2020-10-09 | 2021-01-01 | 上海云绅智能科技有限公司 | 3D target detection method and device, electronic equipment and storage medium |
| CN112396126A (en) * | 2020-12-02 | 2021-02-23 | 中山大学 | Target detection method and system based on detection of main stem and local feature optimization |
| CN112507943A (en) * | 2020-12-18 | 2021-03-16 | 华南理工大学 | Visual positioning navigation method, system and medium based on multitask neural network |
| CN112560935A (en) * | 2020-12-11 | 2021-03-26 | 上海集成电路装备材料产业创新中心有限公司 | Method for improving defect detection performance |
| CN112597897A (en) * | 2020-12-24 | 2021-04-02 | 华南农业大学 | Orchard banana detection method, system, equipment and medium based on deep neural network |
| CN112699914A (en) * | 2020-12-02 | 2021-04-23 | 中山大学 | Target detection method and system based on heterogeneous composite backbone |
| CN113011567A (en) * | 2021-03-31 | 2021-06-22 | 深圳精智达技术股份有限公司 | Training method and device of convolutional neural network model |
| CN113008742A (en) * | 2021-02-23 | 2021-06-22 | 中国农业大学 | Method and system for detecting deposition amount of fog drops |
| CN113177559A (en) * | 2021-04-22 | 2021-07-27 | 重庆兆光科技股份有限公司 | Image recognition method, system, device and medium combining breadth and dense convolutional neural network |
| CN113221908A (en) * | 2021-06-04 | 2021-08-06 | 深圳龙岗智能视听研究院 | Digital identification method and equipment based on deep convolutional neural network |
| CN113269182A (en) * | 2021-04-21 | 2021-08-17 | 山东师范大学 | Target fruit detection method and system based on small-area sensitivity of variant transform |
| CN113326925A (en) * | 2021-06-22 | 2021-08-31 | 华南农业大学 | Density graph regression-based flower amount statistical method, system, equipment and medium for convolutional neural network |
| CN113343918A (en) * | 2021-06-30 | 2021-09-03 | 国网山东省电力公司电力科学研究院 | Power equipment identification method, system, medium and electronic equipment |
| CN113408804A (en) * | 2021-06-24 | 2021-09-17 | 广东电网有限责任公司 | Electricity stealing behavior detection method, system, terminal equipment and storage medium |
| CN113486937A (en) * | 2021-06-28 | 2021-10-08 | 华侨大学 | Solid waste identification data set construction system based on convolutional neural network |
| CN113723217A (en) * | 2021-08-09 | 2021-11-30 | 南京邮电大学 | Object intelligent detection method and system based on yolo improvement |
| CN113723593A (en) * | 2021-08-26 | 2021-11-30 | 国网山东省电力公司济南供电公司 | Load shedding prediction method and system based on neural network |
| CN113807311A (en) * | 2021-09-29 | 2021-12-17 | 中国人民解放军国防科技大学 | Multi-scale target identification method |
| CN113805235A (en) * | 2020-06-11 | 2021-12-17 | 中国石油天然气股份有限公司 | Three-dimensional seismic facies identification method and device based on convolutional neural network |
| CN113850349A (en) * | 2021-11-29 | 2021-12-28 | 苏州优鲜信网络生活服务科技有限公司 | Detection system and method with data identification function |
| CN113920067A (en) * | 2021-09-24 | 2022-01-11 | 电子科技大学 | CT image target detection method, device and equipment based on convolutional neural network |
| CN113945569A (en) * | 2021-09-30 | 2022-01-18 | 河北建投新能源有限公司 | Ion membrane fault detection method and device |
| CN113962035A (en) * | 2021-09-23 | 2022-01-21 | 西安交通大学 | Method and system for predicting pressure between damping surrounding bands of turbine mechanical blade based on convolutional neural network |
| CN114295050A (en) * | 2021-12-24 | 2022-04-08 | 重庆创筑科技有限公司 | Multi-size detection method and system for precast concrete plate based on convolutional neural network |
| CN115147348A (en) * | 2022-05-05 | 2022-10-04 | 合肥工业大学 | Improved YOLOv 3-based tire defect detection method and system |
| CN115630660A (en) * | 2022-12-23 | 2023-01-20 | 湖北凯乐仕通达科技有限公司 | Barcode positioning method and device based on convolutional neural network |
| CN115640835A (en) * | 2022-12-22 | 2023-01-24 | 阿里巴巴（中国）有限公司 | Deep learning network structure generation method and device |
Citations (8)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN105975931A (en) * | 2016-05-04 | 2016-09-28 | 浙江大学 | Convolutional neural network face recognition method based on multi-scale pooling |
| CN106682628A (en) * | 2016-12-30 | 2017-05-17 | 佳都新太科技股份有限公司 | Face attribute classification method based on multilayer depth feature information |
| CN107220618A (en) * | 2017-05-25 | 2017-09-29 | 中国科学院自动化研究所 | Method for detecting human face and device, computer-readable recording medium, equipment |
| CN107844785A (en) * | 2017-12-08 | 2018-03-27 | 浙江捷尚视觉科技股份有限公司 | A kind of method for detecting human face based on size estimation |
| CN108073917A (en) * | 2018-01-24 | 2018-05-25 | 燕山大学 | A kind of face identification method based on convolutional neural networks |
| CN108363962A (en) * | 2018-01-25 | 2018-08-03 | 南京邮电大学 | A kind of method for detecting human face and system based on multi-level features deep learning |
| US20180259608A1 (en) * | 2015-11-29 | 2018-09-13 | Arterys Inc. | Automated cardiac volume segmentation |
| CN109359555A (en) * | 2018-09-21 | 2019-02-19 | 江苏安凰领御科技有限公司 | A kind of high-precision human face quick detection method |
-
2019
- 2019-06-05 CN CN201910483258.4A patent/CN110188720A/en active Pending
Patent Citations (8)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20180259608A1 (en) * | 2015-11-29 | 2018-09-13 | Arterys Inc. | Automated cardiac volume segmentation |
| CN105975931A (en) * | 2016-05-04 | 2016-09-28 | 浙江大学 | Convolutional neural network face recognition method based on multi-scale pooling |
| CN106682628A (en) * | 2016-12-30 | 2017-05-17 | 佳都新太科技股份有限公司 | Face attribute classification method based on multilayer depth feature information |
| CN107220618A (en) * | 2017-05-25 | 2017-09-29 | 中国科学院自动化研究所 | Method for detecting human face and device, computer-readable recording medium, equipment |
| CN107844785A (en) * | 2017-12-08 | 2018-03-27 | 浙江捷尚视觉科技股份有限公司 | A kind of method for detecting human face based on size estimation |
| CN108073917A (en) * | 2018-01-24 | 2018-05-25 | 燕山大学 | A kind of face identification method based on convolutional neural networks |
| CN108363962A (en) * | 2018-01-25 | 2018-08-03 | 南京邮电大学 | A kind of method for detecting human face and system based on multi-level features deep learning |
| CN109359555A (en) * | 2018-09-21 | 2019-02-19 | 江苏安凰领御科技有限公司 | A kind of high-precision human face quick detection method |
Non-Patent Citations (2)
| Title |
|---|
| 公安部第三研究所: "《多摄像机协同关注目标检测跟踪技术》", 30 June 2017 * |
| 张世强: "《中国智能金融产业蓝皮书》", 31 March 2018 * |
Cited By (69)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110569841A (en) * | 2019-09-02 | 2019-12-13 | 中南大学 | contact gateway key component target detection method based on convolutional neural network |
| CN110569841B (en) * | 2019-09-02 | 2022-09-16 | 中南大学 | Contact gateway key component target detection method based on convolutional neural network |
| CN110647622A (en) * | 2019-09-29 | 2020-01-03 | 北京金山安全软件有限公司 | Interactive data validity identification method and device |
| CN111008641B (en) * | 2019-10-24 | 2023-04-07 | 云南电网有限责任公司昆明供电局 | Power transmission line tower external force damage detection method based on convolutional neural network |
| CN111008641A (en) * | 2019-10-24 | 2020-04-14 | 云南电网有限责任公司昆明供电局 | Power transmission line tower external force damage detection method based on convolutional neural network |
| CN110929670A (en) * | 2019-12-02 | 2020-03-27 | 合肥城市云数据中心股份有限公司 | Muck truck cleanliness video identification and analysis method based on yolo3 technology |
| CN111127495A (en) * | 2019-12-04 | 2020-05-08 | 深圳市图郅创新科技有限公司 | Neural network model building method for image key point detection |
| CN111079683A (en) * | 2019-12-24 | 2020-04-28 | 天津大学 | Remote sensing image cloud and snow detection method based on convolutional neural network |
| CN111079683B (en) * | 2019-12-24 | 2023-12-12 | 天津大学 | Remote sensing image cloud and snow detection method based on convolutional neural network |
| CN111191570A (en) * | 2019-12-26 | 2020-05-22 | 国网北京市电力公司 | Image recognition method and device |
| CN111191570B (en) * | 2019-12-26 | 2023-04-18 | 国网北京市电力公司 | Image recognition method and device |
| CN111079861A (en) * | 2019-12-31 | 2020-04-28 | 国网北京市电力公司 | Power distribution network voltage abnormity diagnosis method based on image rapid processing technology |
| CN111222462A (en) * | 2020-01-07 | 2020-06-02 | 河海大学 | Target detection-based intelligent labeling method for apparent feature monitoring data |
| CN111353976A (en) * | 2020-02-25 | 2020-06-30 | 南京大学 | Sand grain target detection method based on convolutional neural network |
| CN111353504B (en) * | 2020-03-02 | 2023-05-26 | 济南大学 | Source camera identification method based on image block diversity selection and residual prediction module |
| CN111353504A (en) * | 2020-03-02 | 2020-06-30 | 济南大学 | Source and machine identification method based on image block diversity selection and residual prediction module |
| CN111523713A (en) * | 2020-04-14 | 2020-08-11 | 中国地质大学（北京） | Method and device for predicting residual oil saturation distribution in oil field |
| CN111523713B (en) * | 2020-04-14 | 2024-01-02 | 中国地质大学（北京） | Method and device for predicting saturation distribution of residual oil in oil field |
| CN111539178A (en) * | 2020-04-26 | 2020-08-14 | 成都市深思创芯科技有限公司 | Chip layout design method and system based on neural network and manufacturing method |
| CN111833395A (en) * | 2020-06-04 | 2020-10-27 | 西安电子科技大学 | Direction-finding system single target positioning method and device based on neural network model |
| CN111833395B (en) * | 2020-06-04 | 2022-11-29 | 西安电子科技大学 | Direction-finding system single target positioning method and device based on neural network model |
| CN113805235B (en) * | 2020-06-11 | 2023-12-26 | 中国石油天然气股份有限公司 | Three-dimensional seismic phase identification method and device based on convolutional neural network |
| CN113805235A (en) * | 2020-06-11 | 2021-12-17 | 中国石油天然气股份有限公司 | Three-dimensional seismic facies identification method and device based on convolutional neural network |
| CN111898598A (en) * | 2020-07-03 | 2020-11-06 | 贵州大学 | Target detection method based on text in dynamic scene |
| CN111898598B (en) * | 2020-07-03 | 2022-08-30 | 贵州大学 | Target detection method based on text in dynamic scene |
| CN112084950A (en) * | 2020-09-10 | 2020-12-15 | 上海庞勃特科技有限公司 | Target detection method and detection device based on sparse convolutional neural network |
| CN112101449B (en) * | 2020-09-11 | 2022-07-15 | 南京邮电大学 | Method for inverting radar data by satellite cloud picture based on semantic loss |
| CN112101449A (en) * | 2020-09-11 | 2020-12-18 | 南京邮电大学 | Method for inverting radar data by satellite cloud picture based on semantic loss |
| CN112098997A (en) * | 2020-09-18 | 2020-12-18 | 欧必翼太赫兹科技(北京)有限公司 | Three-dimensional holographic imaging security inspection radar image foreign matter detection method |
| CN112098997B (en) * | 2020-09-18 | 2021-10-15 | 欧必翼太赫兹科技(北京)有限公司 | Three-dimensional holographic imaging security inspection radar image foreign matter detection method |
| CN112163541A (en) * | 2020-10-09 | 2021-01-01 | 上海云绅智能科技有限公司 | 3D target detection method and device, electronic equipment and storage medium |
| CN112396126B (en) * | 2020-12-02 | 2023-09-22 | 中山大学 | Target detection method and system based on detection trunk and local feature optimization |
| CN112699914A (en) * | 2020-12-02 | 2021-04-23 | 中山大学 | Target detection method and system based on heterogeneous composite backbone |
| CN112699914B (en) * | 2020-12-02 | 2023-09-22 | 中山大学 | Target detection method and system based on heterogeneous composite trunk |
| CN112396126A (en) * | 2020-12-02 | 2021-02-23 | 中山大学 | Target detection method and system based on detection of main stem and local feature optimization |
| CN112560935B (en) * | 2020-12-11 | 2022-04-01 | 上海集成电路装备材料产业创新中心有限公司 | Method for improving defect detection performance |
| CN112560935A (en) * | 2020-12-11 | 2021-03-26 | 上海集成电路装备材料产业创新中心有限公司 | Method for improving defect detection performance |
| CN112507943A (en) * | 2020-12-18 | 2021-03-16 | 华南理工大学 | Visual positioning navigation method, system and medium based on multitask neural network |
| CN112507943B (en) * | 2020-12-18 | 2023-09-29 | 华南理工大学 | Visual positioning navigation method, system and medium based on multitasking neural network |
| CN112597897A (en) * | 2020-12-24 | 2021-04-02 | 华南农业大学 | Orchard banana detection method, system, equipment and medium based on deep neural network |
| CN113008742A (en) * | 2021-02-23 | 2021-06-22 | 中国农业大学 | Method and system for detecting deposition amount of fog drops |
| CN113011567A (en) * | 2021-03-31 | 2021-06-22 | 深圳精智达技术股份有限公司 | Training method and device of convolutional neural network model |
| CN113011567B (en) * | 2021-03-31 | 2023-01-31 | 深圳精智达技术股份有限公司 | Training method and device of convolutional neural network model |
| CN113269182A (en) * | 2021-04-21 | 2021-08-17 | 山东师范大学 | Target fruit detection method and system based on small-area sensitivity of variant transform |
| CN113177559B (en) * | 2021-04-22 | 2023-06-02 | 重庆兆光科技股份有限公司 | Image recognition method, system, equipment and medium combining breadth and dense convolutional neural network |
| CN113177559A (en) * | 2021-04-22 | 2021-07-27 | 重庆兆光科技股份有限公司 | Image recognition method, system, device and medium combining breadth and dense convolutional neural network |
| CN113221908A (en) * | 2021-06-04 | 2021-08-06 | 深圳龙岗智能视听研究院 | Digital identification method and equipment based on deep convolutional neural network |
| CN113221908B (en) * | 2021-06-04 | 2024-04-16 | 深圳龙岗智能视听研究院 | Digital identification method and device based on deep convolutional neural network |
| CN113326925A (en) * | 2021-06-22 | 2021-08-31 | 华南农业大学 | Density graph regression-based flower amount statistical method, system, equipment and medium for convolutional neural network |
| CN113326925B (en) * | 2021-06-22 | 2023-11-28 | 华南农业大学 | Method, system, equipment and medium for calculating flower quantity of convolutional neural network based on density map regression |
| CN113408804A (en) * | 2021-06-24 | 2021-09-17 | 广东电网有限责任公司 | Electricity stealing behavior detection method, system, terminal equipment and storage medium |
| CN113486937A (en) * | 2021-06-28 | 2021-10-08 | 华侨大学 | Solid waste identification data set construction system based on convolutional neural network |
| CN113343918A (en) * | 2021-06-30 | 2021-09-03 | 国网山东省电力公司电力科学研究院 | Power equipment identification method, system, medium and electronic equipment |
| CN113723217A (en) * | 2021-08-09 | 2021-11-30 | 南京邮电大学 | Object intelligent detection method and system based on yolo improvement |
| CN113723593A (en) * | 2021-08-26 | 2021-11-30 | 国网山东省电力公司济南供电公司 | Load shedding prediction method and system based on neural network |
| CN113723593B (en) * | 2021-08-26 | 2024-01-09 | 国网山东省电力公司济南供电公司 | Cut load prediction method and system based on neural network |
| CN113962035B (en) * | 2021-09-23 | 2024-04-05 | 西安交通大学 | Turbine mechanical blade damping inter-shroud pressure prediction method and system based on convolutional neural network |
| CN113962035A (en) * | 2021-09-23 | 2022-01-21 | 西安交通大学 | Method and system for predicting pressure between damping surrounding bands of turbine mechanical blade based on convolutional neural network |
| CN113920067A (en) * | 2021-09-24 | 2022-01-11 | 电子科技大学 | CT image target detection method, device and equipment based on convolutional neural network |
| CN113807311A (en) * | 2021-09-29 | 2021-12-17 | 中国人民解放军国防科技大学 | Multi-scale target identification method |
| CN113945569B (en) * | 2021-09-30 | 2023-12-26 | 河北建投新能源有限公司 | Fault detection method and device for ion membrane |
| CN113945569A (en) * | 2021-09-30 | 2022-01-18 | 河北建投新能源有限公司 | Ion membrane fault detection method and device |
| CN113850349A (en) * | 2021-11-29 | 2021-12-28 | 苏州优鲜信网络生活服务科技有限公司 | Detection system and method with data identification function |
| CN114295050A (en) * | 2021-12-24 | 2022-04-08 | 重庆创筑科技有限公司 | Multi-size detection method and system for precast concrete plate based on convolutional neural network |
| CN114295050B (en) * | 2021-12-24 | 2024-05-17 | 重庆创筑科技有限公司 | Precast concrete slab multi-size detection method and system based on convolutional neural network |
| CN115147348A (en) * | 2022-05-05 | 2022-10-04 | 合肥工业大学 | Improved YOLOv 3-based tire defect detection method and system |
| CN115640835B (en) * | 2022-12-22 | 2023-03-31 | 阿里巴巴（中国）有限公司 | Deep learning network structure generation method and device |
| CN115640835A (en) * | 2022-12-22 | 2023-01-24 | 阿里巴巴（中国）有限公司 | Deep learning network structure generation method and device |
| CN115630660A (en) * | 2022-12-23 | 2023-01-20 | 湖北凯乐仕通达科技有限公司 | Barcode positioning method and device based on convolutional neural network |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN110188720A (en) | A kind of object detection method and system based on convolutional neural networks | |
| CN106780612B (en) | Object detecting method and device in a kind of image | |
| CN107145908B (en) | A kind of small target detecting method based on R-FCN | |
| CN109101930A (en) | A kind of people counting method and system | |
| CN107273836A (en) | A kind of pedestrian detection recognition methods, device, model and medium | |
| CN106874894A (en) | A kind of human body target detection method based on the full convolutional neural networks in region | |
| CN110163889A (en) | Method for tracking target, target tracker, target following equipment | |
| CN107967451A (en) | A kind of method for carrying out crowd's counting to static image using multiple dimensioned multitask convolutional neural networks | |
| CN110084173A (en) | Number of people detection method and device | |
| CN109446925A (en) | A kind of electric device maintenance algorithm based on convolutional neural networks | |
| CN108520229A (en) | Image detecting method, device, electronic equipment and computer-readable medium | |
| CN109271888A (en) | Personal identification method, device, electronic equipment based on gait | |
| CN109800628A (en) | A kind of network structure and detection method for reinforcing SSD Small object pedestrian detection performance | |
| CN109934115A (en) | Construction method, face identification method and the electronic equipment of human face recognition model | |
| CN110084195A (en) | Remote Sensing Target detection method based on convolutional neural networks | |
| CN110363104A (en) | A kind of detection method of diesel oil black smoke vehicle | |
| CN104850836A (en) | Automatic insect image identification method based on depth convolutional neural network | |
| CN110348441A (en) | VAT invoice recognition methods, device, computer equipment and storage medium | |
| CN109598234A (en) | Critical point detection method and apparatus | |
| CN109558902A (en) | A kind of fast target detection method | |
| CN110276264A (en) | A kind of crowd density estimation method based on foreground segmentation figure | |
| CN109740676A (en) | Object detection moving method based on similar purpose | |
| CN106372597B (en) | CNN Vehicle Detection method based on adaptive contextual information | |
| CN108960015A (en) | A kind of vehicle system automatic identifying method and device based on deep learning | |
| CN107092883A (en) | Object identification method for tracing |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| RJ01 | Rejection of invention patent application after publication |
Application publication date: 20190830 |
|
| RJ01 | Rejection of invention patent application after publication |