CN109615010A - Chinese traditional medicinal materials recognition method and system based on double scale convolutional neural networks - Google Patents
Chinese traditional medicinal materials recognition method and system based on double scale convolutional neural networks Download PDFInfo
- Publication number
- CN109615010A CN109615010A CN201811525966.1A CN201811525966A CN109615010A CN 109615010 A CN109615010 A CN 109615010A CN 201811525966 A CN201811525966 A CN 201811525966A CN 109615010 A CN109615010 A CN 109615010A
- Authority
- CN
- China
- Prior art keywords
- medicinal material
- image
- layer
- convolutional neural
- neural networks
- 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.)
- Granted
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
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/25—Fusion techniques
- G06F18/253—Fusion techniques of extracted features
Landscapes
- Engineering & Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Theoretical Computer Science (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Artificial Intelligence (AREA)
- Evolutionary Biology (AREA)
- Evolutionary Computation (AREA)
- Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Image Analysis (AREA)
Abstract
This application discloses the Chinese traditional medicinal materials recognition method and system based on double scale convolutional neural networks, the acquisition of double scale images is carried out by the image collecting device specially designed first, training image feeding convolutional neural networks are trained, by the feature extraction and selection in multilayer convolution pond, training obtains the higher convolutional neural networks model of accuracy of identification.Testing image is sent into trained convolutional neural networks model and carries out feature extraction, the feature based on extraction classifies to Chinese medicine, output category recognition result.The present invention is conducive to improve ordinary people to the recognition capability of medicinal material, assists the more acurrate quick discrimination medicinal material of medicinal material domain expert.
Description
Technical field
This disclosure relates to the Chinese traditional medicinal materials recognition method and system based on double scale convolutional neural networks.
Background technique
The statement of this part is only to improve background technique relevant to the disclosure, not necessarily constitutes the prior art.
With the growth of China's economic level, national living standard is also higher and higher.In order to improve the quality of living, more
People start focus on health.Since ancient times, TCM Culture is of extensive knowledge and profound scholarship, contains the history essence of China over the past thousands of years, it
It is the Chinese nation and the treasure that disease long-term struggle accumulates in the process, is the important composition portion of Chinese nation's excellent culture
Point, the prosperity for human health and the Chinese nation is made that great contribution.
As the conventional industries in China, Chinese Medicine Industry has in terms of resource and knowledge and has great advantage.Inherit China
Traditional medicine culture, makes further innovation on this basis, pushes the development of China's medicinal industry.Chinese medicine is as a kind of biography
System drug, not only plays an important role traditional chinese medicine, but also also have significant impact for modern medicine, has found
Method is come the problem for the treatment of type-II diabetes, cancer and female infertility etc..
Chinese medicine it is proper use of, play the role of to the life security of patient vital.Present Chinese medicine field
Expert is rare, and many Pharmacy practice doctors are unfamiliar with the discrimination of medicinal material, and mistake uses medicinal material, it is most likely that causes very serious
Consequence.The true medicine of Market of Chinese Materia Medica mixes counterfeit drug at present, even medical related personnel also can not necessarily ensure the positive confirmation of medicinal material
Know.Chinese medicine huge number greatly increases the difficulty of discrimination.
Application No. is 2017109396722, a kind of denomination of invention are as follows: Chinese traditional medicinal materials recognition side based on deep neural network
Method, uses web crawlers and manually takes pictures and acquire Chinese medicine picture as the input of data set and pre-processed;Training with
Prediction process uses the Bagging method of integrated study, is carried out using classical convolutional neural networks model to each sub- training set
Fine tuning training, generates multiple Weak Classifiers, finally cooperates Softmax sorting algorithm, and obtain using integrated study combined strategy
Strong classifier obtains classification results.Although the patent can be realized Chinese traditional medicinal materials recognition, still, it is applicant's understanding that the people of the patent
Work link of taking pictures is not taken pictures using special device, not can avoid insufficient light of taking pictures, the reflective, hand shaking of taking pictures of taking pictures the problems such as bat
The image that cannot accurately identify medicinal material type come is taken out, it is not high to will lead to later period convolutional neural networks accuracy of identification;And this is specially
Benefit is using common convolutional neural networks algorithm, the problem is that recognition speed is unhappy.
Summary of the invention
In order to solve the deficiencies in the prior art, present disclose provides the Chinese traditional medicinal materials recognitions based on double scale convolutional neural networks
Method and system;
In a first aspect, present disclose provides the Chinese traditional medicinal materials recognition methods based on double scale convolutional neural networks;
Chinese traditional medicinal materials recognition method based on double scale convolutional neural networks, comprising:
Double scale convolutional neural networks are constructed, acquire the whole drawn game under every kind of medicinal material equal resolution using drawer box
Training image is opened in portion two, and the corresponding classification of medicinal material label of setting training image carries out two training images of every kind of medicinal material
Pretreatment, regard two training images of pretreated every kind of medicinal material as one group of input value side by side, is input to convolutional Neural net
In network, using classification of medicinal material label as the output valve of convolutional neural networks, convolutional neural networks are trained, are trained
Convolutional neural networks;
In the same way, the entirety under the equal resolution of medicinal material to be identified and part two are acquired using drawer box
Testing image is opened, two testing images of medicinal material to be identified are pre-processed, regard two testing images as one group side by side
Input value is input in trained convolutional neural networks, extracts two to mapping using trained convolutional neural networks
It is after the multi-features of picture as a result, obtaining the identification knot of the corresponding medicinal material of the testing image according to fused result
Fruit.
As a kind of possible implementation, double scale convolutional neural networks, comprising:
Sequentially connected input layer, the first convolutional layer, the first pond layer, the second convolutional layer, the second pond layer, third convolution
Layer, third pond layer, full articulamentum and output layer;
The input layer, including six channels, wherein three channels are used to receive the general image of the first scale, in addition three
A channel is used to receive the topography of the second scale;Every image includes tri- channels RGB i.e. 3 picture element matrixs;It is described defeated
Enter the input of layer for six picture element matrixs, in six picture element matrixs, there are three picture element matrixs from the general image of shooting, separately
The topography from shooting of outer three picture element matrixs;Matrix size is 92*92*6;
First convolutional layer extracts characteristics of image, volume for carrying out convolution operation with the picture element matrix of the input layer
Product core selects 5*5*6*20, and biasing size is 20*1, show that characteristic pattern size is 88*88*20;
First pond layer, the characteristics of image picture element matrix for extracting to first convolutional layer carry out dimensionality reduction behaviour
Make, extract notable feature information, selection convolution kernel size is 2*2, and the characteristic pattern size after dimensionality reduction is 44*44*20;
Second convolutional layer, the characteristic pattern for obtaining to first pond layer further carry out convolution operation,
Characteristics of image is extracted, convolution kernel selects 5*5*6*24, and biasing size is 24*1, show that characteristic pattern size is 40*40*24;
Second pond layer, the further dimensionality reduction of characteristic pattern for obtaining to second convolutional layer extract significant special
Reference breath, selection convolution kernel size are 2*2, and the characteristic pattern size after dimensionality reduction is 20*20*24;
The third convolutional layer, the further convolution operation of characteristic pattern for obtaining to second pond layer extract figure
As feature, convolution kernel selects 5*5*6*20, and biasing size is 24*1, show that characteristic pattern size is 16*16*24;
Third pond layer, the further dimensionality reduction operation of characteristic pattern for obtaining to the third convolutional layer, is extracted aobvious
Characteristic information is write, selection convolution kernel size is 2*2, and characteristic pattern size is 8*8*24 after dimensionality reduction;
The full articulamentum, the characteristics of image for obtaining to third pond layer are integrated, weight matrix size
For p*1536, it is biased to p*1.
The output layer, for exporting identification classification results, output matrix size is p*1, which represents input sample
In the corresponding other probability of p kind herbs Class, be maximized where classification be classification results.
Two images of the entirety of same Chinese medicine and part under equal resolution are why acquired, rather than are simply schemed
As scaling, the small Chinese medicine of particle or the similar Chinese medicine of texture are allowed for;The small Chinese medicine of the particle such as Semen Cuscutae
And semen astragali complanati;The similar Chinese medicine of the texture such as root bark of tree peony and cortex dictamni;The conventional picture shot later that focuses is seen simultaneously on the whole
It is that can not accurately extract its high-rise spy if this training sample is directly sent in convolutional neural networks without significant difference
Sign.Therefore the application acquires the training image of two kinds of scales to every a medicinal material, just helps since the extraction of low-level image feature
Convolutional neural networks obtain the low-level details such as the more textures of training image, shape, color as far as possible, improve training effectiveness and most
Whole accuracy of identification.
As a kind of possible implementation, for the specific of two different training images of the equal shooting angle of every kind of medicinal material
Step are as follows:
Each medicinal material is shot using drawer box, the drawer box, including square box ontology, the side
Shape cassette body includes top surface, bottom surface, left side, right side, leading flank and trailing flank, the inside of the left side and right side
It is equipped with sliding rail, the sliding rail is separately positioned on the middle position of the vertical direction of left side and right side;The square box
Body interior middle position is equipped with drawer, the sliding rail being separately mounted on the inside of left side and right side at left and right sides of the drawer
On;Square box ontology is divided into two spaces up and down by the drawer, and the square box top surface is equipped with rectangular slot, described
Rectangular slot is for placing mobile phone;The mobile phone is connect with background server;
Before shooting, medicinal material is put into drawer, after closing drawer, starts the photographing mode of mobile phone, to the medicinal material in drawer
Shoot first training image;
Then, drawer is extracted out from square box ontology, medicinal material is placed on square box bottom surface, mobile phone is started
The medicinal material of photographing mode, square shaped box bottom surface shoots second training image;
First training image is the topography for showing medicinal material regional area, and second training image is that display medicinal material is whole
The general image of body region;Double scale images refer to that the image source data of input is entirety and part under equal resolution
The image in two kinds of visuals field.
First training image and second training image are uploaded to background server simultaneously.
The drawer box, when drawer is closed, entire drawer is sealing, due to there is about one outside drawer
Plug-type door, so entire cassette interior is still sealing after drawer is taken away, as medicinal material establishes a camera bellows environment, from
And ambient light is avoided to enter drawer cassette interior, and then from interfering the shooting of medicinal material.
The drawer box, top surface are equipped with the lamp bead of S-shaped arrangement, and the lamp bead is powered by battery, and the lamp bead issues
Light be collimated white light, to guarantee that drawer box internal brightness is consistent.Selecting white light is that the light of other colors in order to prevent shines
It is mapped on medicinal material, the color of medicinal material itself may be changed and influence recognition effect;Select directional light rather than point light source, be because
For directional light compared to point light source, intensity of illumination everywhere is more balanced, is not likely to produce light shift and hot spot, it is more difficult to generate bright
Aobvious inclined dark areas can preferably guarantee the authenticity and validity of clapped medicinal material picture.
The drawer box uses the resin material of white, and the opaque paster of one layer of black is posted outside drawer box,
It avoids light from being irradiated to drawer cassette interior, generates reflective influence to taking pictures, guarantee the equipment in addition to internal light source, Bu Huizai
Receive the influence of other any light sources.
The mobile phone is placed on rectangular slot, and when can be to avoid taking pictures, the shooting angle of people or hand-motion jitter etc. be not really
Determine the influence that factor generates the clarity of image.
As a kind of possible implementation, the specific steps of the corresponding classification of medicinal material label of training image are set are as follows:
Medicinal material unique number and medicinal material corresponding title are set on tag along sort.
As a kind of possible implementation, pretreated specific steps are carried out to two training images of every kind of medicinal material
Are as follows:
For the quantity of spread training image, random process carried out to the image of shooting, and to obtaining after random process
Training image before image and processing is together as training image;The random process, comprising: Random-Rotation or overturning;At random
Tone reversal, Random-fuzzy or random light shift field.
As a kind of possible implementation, it regard two training images of pretreated every kind of medicinal material as one group side by side
Input value is input in convolutional neural networks, using classification of medicinal material label as the output valve of convolutional neural networks, to convolutional Neural
Network is trained, and obtains the specific steps of trained convolutional neural networks are as follows:
One group of training image of input, including two different training images of shooting angle, every training image includes three
A input matrix;It includes six input matrixes that two training images, which have altogether,;
The one group of image and convolution kernel of input carry out convolution operation, and 1 input matrix corresponds to m convolution kernel and grasps by convolution
Make to show that m characteristic pattern, 6 input matrixes correspond to 6*m convolution kernel, obtains 6*m characteristic pattern by convolution operation, it will
In the 6*m characteristic pattern, corresponding every 6 characteristic patterns carry out Fusion Features, that is, add up to 1 characteristic pattern, that is, finally obtain m
A characteristic pattern;
Pond layer takes average operation to gained characteristic pattern, m picture pixels matrix after final output Fusion Features, matrix
By the Classification and Identification of full articulamentum, the identification probability of p kind medicinal material is exported, according to output result compared with actual result, is obtained
Output layer error out is successively fed back, each layer parameter of regulating networks with gradient descent method by error back propagation;
Repetitive exercise several times is carried out, parameter is adjusted, when the loss function value of convolutional neural networks is less than setting
When threshold value or the number of iterations reach setting number, deconditioning obtains trained convolutional neural networks.
M is a parameter of convolution kernel, and value is 20 or 24.
As a kind of possible implementation, pretreated specific step is carried out to two testing images of medicinal material to be identified
Suddenly are as follows:
Compression processing is carried out to the size of testing image.
As a kind of possible implementation, the figure of two testing images is extracted using trained convolutional neural networks
As the specific steps of the result after Fusion Features are as follows:
S11:1 input matrix of step corresponds to m convolution kernel and obtains m characteristic pattern, and one group of testing image is 6 input squares
The i.e. corresponding 6*m convolution kernel of battle array obtains 6*m characteristic pattern, and the corresponding every 6 progress Fusion Features of characteristic pattern are added up to 1 spy
Sign figure, finally obtains m characteristic pattern;Pond layer takes average operation to gained characteristic pattern, m figure after final output Fusion Features
Piece picture element matrix;
Step S12: by m picture pixels matrix after Fusion Features plus biasing, and then by each member of gained matrix
Element is input in sigmoid activation primitive, extracts the profile and profile each section position feature of image;Subsequently into pond
Layer carries out the dimensionality reduction of picture element matrix, prominent main feature;
Step S13: by the feature extraction and selection of all convolutional layers and pond layer, the fusion of feature is realized;Finally will
Gained eigenmatrix pulls into a column vector in sequence.
As a kind of possible implementation, the knowledge of the corresponding medicinal material of the testing image is obtained according to fused result
The specific steps of other result are as follows:
Step S21: pulling into a column vector for gained eigenmatrix, is sent into full articulamentum, and the column vector is 1536 dimensions;
Step S22: it by column vector multiplication corresponding with each row element of weight matrix, then is added, show that matrix is big
Small is p*1, and weight matrix size is p*1536;
Step S23: it by gained matrix plus biasing, and then is sent into sigmoid activation primitive;
Step S24: by activation primitive output p as a result, representing the other probability of p kind herbs Class, where being maximized
Classification is final result.
Second aspect, present disclose provides the Chinese traditional medicinal materials recognition systems based on double scale multi-features;
Chinese traditional medicinal materials recognition system based on double scale multi-features, comprising: mobile phone includes memory and place in mobile phone
Device is managed, computer instruction is stored in the memory, the processor refers to for handling computer stored in memory
It enables, is completed at the same time step described in any of the above-described method.
The third aspect, present disclose provides a kind of computer readable storage mediums;
A kind of computer readable storage medium, operation has computer instruction thereon, and the computer instruction is transported by processor
When row, step described in any of the above-described method is completed.
Compared with prior art, the beneficial effect of the disclosure is:
1, this application provides a kind of Novel drawer box for acquisition, being acquired using this device effectively to be kept away
The influence that the uncertain factors such as light conditions, shooting angle and hand trembling generate when exempting from artificially to shoot, improves system identification
When robustness and stability.
2, using double inputs of the scale image as convolutional neural networks, i.e., same two kinds of scalograms of medicinal material in the application
Piece is fed together convolutional neural networks, and convolutional neural networks can be helped to extract the height of the very much like medicinal material of some textural shapes
Layer distinguishing characteristics, the effective accuracy of identification for improving convolutional neural networks.It can adapt to people in reality using double scale images
The problem of different angle is shot, is allowed to can still provide for accurately identifying in different visual field shooting process.It is same using the input of double scales
When can also obtain the detailed information such as texture, the shape of different scale, improve the discrimination of neural network.Double scale convolutional Neurals
Network inputs two images simultaneously, is also beneficial to the quick identification to image, improves the speed of image recognition.
3, the present invention is a kind of real-time identifying system, can quickly return to identification knot for medicinal material picture captured by user
Fruit, real-time are preferable.
4, the scarcity of medicinal material field professional person is compensated for, amateur is to the understanding of medicinal material for raising, can be quick and precisely
Medicinal material is recognized.The application effectively compensates for China and greatly develops under Chinese Medicine Industry background, and medicinal material identifies personnel not
Foot, medicinal material identification have difficulties, and similar medicinal material holds confusing deficiency.
Detailed description of the invention
The accompanying drawings constituting a part of this application is used to provide further understanding of the present application, and the application's shows
Meaning property embodiment and its explanation are not constituted an undue limitation on the present application for explaining the application.
Fig. 1 is the step flow chart of the application;
Fig. 2 is the drawer box schematic diagram of the application;
Fig. 3 is the structural schematic diagram of the application convolutional neural networks;
Fig. 4 is the structural schematic diagram of the application convolutional neural networks.
Specific embodiment
It is noted that described further below be all exemplary, it is intended to provide further instruction to the application.Unless another
It indicates, all technical and scientific terms used herein has usual with the application person of an ordinary skill in the technical field
The identical meanings of understanding.
It should be noted that term used herein above is merely to describe specific embodiment, and be not intended to restricted root
According to the illustrative embodiments of the application.As used herein, unless the context clearly indicates otherwise, otherwise singular
Also it is intended to include plural form, additionally, it should be understood that, when in the present specification using term "comprising" and/or " packet
Include " when, indicate existing characteristics, step, operation, device, component and/or their combination.
Classified using double scale medicinal material picture samples of the convolutional neural networks to input.Due to the medicinal material being sent into simultaneously
Picture is whole and two pictures of part, and a picture includes that tri- channels RGB are 3 dimension matrixes, i.e., convolutional neural networks is defeated
Enter for six channels, 6 matrixes.It is successively fed back, is instructed according to the output result of convolutional neural networks and the difference of practical medicinal material information
Each parameter value for practicing neural network, finally obtains accurate classification results, as shown in Figure 4.
Embodiment one
As shown in Figure 1, the Chinese traditional medicinal materials recognition method based on double scale multi-features, includes the following steps:
Step S1: acquisition medicinal material picture pre-processes picture.In order to improve the accuracy rate of machine learning, acquiring
During medicinal material, the same angle shot is clapped twice for same medicinal material, and one remote one is close, and dimension of picture is unified for 500*
Then the far and near picture of same medicinal material is sent into picture processing function by 500 sizes, to picture by random cropping, random rotation
Turn or the processing such as overturning, arbitrary color tone transformation, Random-fuzzy, random light shift field after Output Size for 92*92 size figure
Input of the piece as convolutional neural networks model.
Step S2: since data set is excessively huge, random each 5 pictures of choosing are instructed from far and near pictures every time
Practice, it is ensured that concentrating 5 pictures in 5 pictures chosen and nearly pictures from long-term plan piece is one-to-one relationship.Every figure
The picture element matrix size of piece is 92*92*3, and two pictures of distance are sent into simultaneously, i.e., input size is 92*92*6, by
Triple channel before becomes six channels, is sent into convolutional neural networks and is trained.
Step S3: as shown in figure 3, convolutional neural networks take three-layer coil lamination and three layers of pond layer, the first convolutional layer volume
Product core is 5*5*6*20, exports 20 characteristic patterns;Second convolutional layer convolution kernel is 5*5*6*24, exports 24 characteristic patterns;Third
Convolutional layer convolution kernel is 5*5*6*24, exports 24 characteristic patterns.This three layers pond layer convolution kernel is 2*2.
Step S4: initialization convolutional neural networks, each parameter tax are provided that learning rate 0.1, and the number of iterations 3000 is defeated
Enter 2 pictures i.e. 6 matrix, picture size 92*92, convolution kernel size is 5*5, for the first, second and third convolutional layer, convolution kernel
Assignment range is the random number between -1 to 1, and biasing initial value is 0, for the first, second and third pond layer, the Chi Huafang of use
Formula is that mean-pooling be averaged pond mode, that is, taking convolution kernel value is 0.25, and biasing initial value is 0, the power of full articulamentum
Weight values take the random number between -1 to 1, and biasing initial value is assigned to 0.
Step S5: each small training set is trained using convolutional neural networks, in convolution process, for the first volume
Lamination, 1 input matrix correspond to m convolution kernel and show that m characteristic pattern, 6 input matrixes correspond to 6*m by convolution operation
Convolution kernel obtains 6*m characteristic pattern by convolution operation, this 6*m characteristic pattern is added up to every 6 progress Fusion Features
1 characteristic pattern, that is, finally obtain m characteristic pattern.M is a parameter of convolution kernel, it is proposed that takes 20.For it is subsequent second and third
Regular volume lamination, the number of the characteristic pattern of output are equal to the number of the input feature vector figure of this layer of convolution.During this, the
One, formula used in two, three convolutional layers is as follows:
Wherein a is output matrix, and I is double scale picture matrixes of input, and ω is convolution kernel, and b is biasing, and f refers to swashing
Function living, the application use sigmoid activation primitive.Function formula is as follows:
Pond layer is using average pond mode, mean-pooling method.
Step S6: 24 characteristic patterns are exported by third pond layer, size 4*4 pulls into these picture pixels matrix values
One column vector is as full articulamentum, and 8*8*24=1536 neuron, the weight matrix size of this layer are 17*1536 in total, partially
It is set to 17*1, the adduction by the multiplication corresponding with weight of each neuron is obtained along with biasing by sigmoid activation primitive
17 kinds of other probability of herbs Class.
Step S7: according to the result obtained compared with actual result, reversely being adjusted the parameter of convolutional neural networks,
The result for obtaining it is more accurate.Here the loss function used comes assessment result, MSE for classical mean square error (MSE)
Smaller, the modelling effect of Classification of Chinese Drug is better.
Wherein n is class categories, and act is sample actual value, and pre is predicted value, and belonging to such predicted value is 1, is not belonging to
Such predicted value is 0.
Step S8: calculating the error between output layer result and actual result, and the error that backpropagation is returned can be seen
Work is the sensitivity of each neuron.The neuron sensitivity of output layer are as follows:
uL=ωLxL-1+bL
δ indicates sensitivity, the derivative of f ' expression activation primitive, and y indicates output valve, and t indicates actual value.
Representing matrix corresponding element is multiplied.ω indicates that weight, x indicate upper one layer of input, and b indicates biasing.
Step S9: the error of full articulamentum is the input and this layer of sensitivity of this layer for the derivative of this layer each weight
Multiplication cross, obtained partial derivative is exactly the right value update of this layer of neuron multiplied by learning rate, and formula is as follows:
Wherein E indicates mean square error, and δ indicates sensitivity, that is, error, and f ' indicates the derivative of activation primitive,Representing matrix pair
Answer element multiplication.ω indicates that weight, x indicate upper one layer of input, and η indicates learning rate.
Step S10: being obtained the error of full articulamentum by step S9, since full articulamentum is the column that third pond layer pulls into
Vector, so the error of full articulamentum passes through matrix reconstruction, that is, third pond layer error.
Step S11: the third pond layer error as known to step S10 continues reverse-direction derivation according to this layer of error, after diminution
The error in region restores the error of preceding layer large area, this process is called upsample.Third pond layer region size
For 2*2, error matrix is extended each 1 row 1 up and down and arranged, that is, become 4*4 size, the matrix after reduction includes four 2*2
Minor matrix carries out average computation for each minor matrix, and obtained matrix is exactly the matrix after upsample.By third
Pond layer error derives the formula of third convolutional layer error are as follows:
Step S12: when being the first, second, third convolutional layer for the layer, just there is convolution operation in when propagated forward,
Known first, second and third convolutional layer error, when seeking the first and second pond layer error, convolution kernel will overturn 180 ° and be calculated, formula
It is as follows:
Indicate convolution operation, rot180 (x) indicates that x overturns 180 ° up and down
Step S13: calculating the error of the first, second and third convolutional layer, convolution kernel to the first, second and third convolutional layer and partially
It sets to be updated and seeks gradient, newer is as follows:
A indicates one layer of output, after the update i.e. error of this layer of convolution kernel and 180 ° of preceding layer output switching activity
Convolution is done, the update for biasing is that the item of each submatrix of error delta is summed respectively, an error vector is obtained, as b's
Gradient.
Step S14: by successive ignition training, constantly parameter is adjusted, when error is less than 0.1, that is, is terminated
Training.If the minimum value of setting is not achieved in error, i.e. 3000 end of iteration are trained.
Step S15: finally obtaining the classification results of the medicinal material, i.e., defeated for the picture pixels matrix of each 92*92*6
17 results out represent 17 kinds of other probability of herbs Class, and classification where being maximized is the classification of the image.
Embodiment two
The present embodiment specifically acquires the preparation of equipment, training pattern from medicinal material, and the several aspects of actual test introduce a kind of base
In the Chinese traditional medicinal materials recognition technology that double scale multi-features and data enhance, detailed process is as follows:
One, the preparation of medicinal material acquisition equipment
When it is artificial take pictures collect the medicinal material material to be identified when, light conditions when people take pictures, shooting
The uncertain factors such as angle and hand trembling can have an impact the photo of acquirement.Therefore, if not improving this case, i.e.,
The acquired picture effect that makes repeatedly to take pictures for same medicinal material be likely to be it is different, will greatly affect medicinal material
The accuracy rate of identification.So in order to solve this problem, designing acquisition medicinal material equipment using 3D printing technique.
As shown in Fig. 2, being shot using drawer box to each medicinal material, the drawer box, including square box
Ontology, the square box ontology include top surface, bottom surface, left side, right side, leading flank and trailing flank, the left side and
Sliding rail is equipped on the inside of right side, the sliding rail is separately positioned on the middle position of the vertical direction of left side and right side;
Square box body interior middle position is equipped with drawer, is separately mounted to left side and right side at left and right sides of the drawer
On sliding rail on the inside of face;Square box ontology is divided into two spaces up and down by the drawer, and the square box top surface is equipped with
Rectangular slot, the rectangular slot is for placing mobile phone;The mobile phone is connect with background server;
Before shooting, medicinal material is put into drawer, after closing drawer, starts the photographing mode of mobile phone, to the medicinal material in drawer
Shoot first training image;
Then, drawer is extracted out from square box ontology, medicinal material is placed on square box bottom surface, mobile phone is started
The medicinal material of photographing mode, square shaped box bottom surface shoots second training image;
First training image is the topography for showing medicinal material regional area, and second training image is that display medicinal material is whole
The general image of body region;Double scale images refer to that the image source data of input is entirety and part under equal resolution
The image in two kinds of visuals field.
First training image and second training image are uploaded to background server simultaneously.
In order to avoid illumination is for the influence taken pictures, equipment is configured to a closed darkroom.In order to avoid light photograph
It is mapped to cabinet interior, generates reflective influence to taking pictures, in the portability for considering equipment, this is printed using the resin material of white
One equipment.Since the color of equipment is white, so extraneous light source can be injected inside equipment, therefore one is sticked outside chest
The lighttight paster of layer black guarantees that the chest in addition to internal light source, will not receive the influence of other any light sources again.
Medicinal material equipment structure chart is shown in attached drawing 1, and equipment specific production step is as follows:
Step S101: the equipment consists of three parts, respectively equipment body, appliance doors, equipment internal drawer part.
Step S102: the 171mm high 35mm 284mm wide long drawer central part in equipment drawer part 104 is slightly recessed, and guarantees
Place the safety of medicinal material.
Step S103: equipment door section 102 is 25mm high 173mm 360mm wide long, reserved in front of equipment body for being placed on
In groove out, the gap of equipment is shut, equipment is made to become a lighttight cuboid.
Step S104: equipment body part is divided into four parts: medicinal material placement part, power supply unit placement part, and mobile phone is put
Set part and the gate slot for being partially placed into for appliance doors.
Step S105: for mobile phone placement part 101 at top, front end is that the roof cuboid of hollow out is clapped to fix mobile phone etc.
It takes pictures according to equipment.
Step S106: power supply unit placement part 103 is 57mm high 75mm 175mm wide long, for placing the power supply such as charger baby
Equipment.
Step S107: equipment body internal partial wall two sides have two attachment straps of long 160mm wide 7.5mm high 5mm to be used to fix
Drawer 104.
Step S108: equipment inside top is provided with the 5050 model lamp bead of 2m long that eight line up serpentine arrangement and is used to guard box
Light source in son is collimated white light.White light purpose is selected, is in order to prevent on the light-illuminating to medicinal material of other colors, it may
Change the color of itself and influences recognition effect.And why select directional light rather than point light source, it is because of directional light phase
Compared with point light source, intensity of illumination everywhere is more balanced, is not likely to produce light shift and hot spot, it is more difficult to generate apparent partially dark
Region can preferably guarantee taken a picture authenticity and validity.
Two, training pattern
Step S201: reading medicinal material picture from the remote and file of shooting at close range respectively, makes tag along sort,
It is saved as .mat form.
Step S202: data file load is come in, and matrix size is processed into 92*92, is then fed into convolutional Neural net
Network.
Step S203: convolutional neural networks are inputted using 6 channels, and two pictures of a medicinal material are sent into net simultaneously
Network carries out feature extraction by multiple convolution ponds layer of convolutional neural networks, the spy after the two picture Fusion Features obtained
Sign figure.
Step S204: the last layer pond layer is pulled into a column vector as full articulamentum.
Step S205: obtaining a result by classification, and result is compared with actual value, calculates error.
Step S206: carrying out error back propagation, used gradient descent method, calculates layer-by-layer error, adjusts network parameter.
Step S207: repeating step S203-step S206, until error less than 0.1 or iteration 3000 times, terminates
Training.
Three, actual test
Step S301: from mobile phone terminal shooting, medicinal material, first bat one are kept in simply, and previous field range is different again therewith
One is clapped, two pictures are sent into background server simultaneously.
Step S302: two different picture sizes of field range are processed into 92*92.
Step S303: it by the picture handled well while being sent into the good convolutional neural networks model of precondition.
Step S304: by the feature extraction and selection of convolutional neural networks convolutional layer pond layer, and then by connecting entirely
Layer Classification and Identification, last output layer export recognition result, and accuracy can achieve 90% or more.
The foregoing is merely preferred embodiment of the present application, are not intended to limit this application, for the skill of this field
For art personnel, various changes and changes are possible in this application.Within the spirit and principles of this application, made any to repair
Change, equivalent replacement, improvement etc., should be included within the scope of protection of this application.
Claims (10)
1. the Chinese traditional medicinal materials recognition method based on double scale convolutional neural networks, characterized in that include:
Double scale convolutional neural networks are constructed, entirety and the training figure of part two under equal resolution are acquired using drawer box
Picture, the corresponding classification of medicinal material label of setting training image, pre-processes two training images of every kind of medicinal material, will locate in advance
Two training images of every kind of medicinal material after reason are used as one group of input value side by side, are input in convolutional neural networks, by medicinal material point
Output valve of the class label as convolutional neural networks, is trained convolutional neural networks, obtains trained convolutional Neural net
Network;
In the same way, acquired using drawer box entirety under the equal resolution of medicinal material to be identified and part two to
Altimetric image, pre-processes two testing images of medicinal material to be identified, regard two testing images as one group of input side by side
Value, is input in trained convolutional neural networks, extracts two testing images using trained convolutional neural networks
It is after multi-features as a result, obtaining the recognition result of the corresponding medicinal material of the testing image according to fused result.
2. the method as described in claim 1, characterized in that described to construct double scale convolutional neural networks, comprising:
Sequentially connected input layer, the first convolutional layer, the first pond layer, the second convolutional layer, the second pond layer, third convolutional layer,
Third pond layer, full articulamentum and output layer;
The input layer, including six channels, wherein three channels are used to receive the general image of the first scale, the other three is logical
Road is used to receive the topography of the second scale;Every image includes tri- channels RGB i.e. 3 picture element matrixs;The input layer
Input be six picture element matrixs, in six picture element matrixs, there are three picture element matrixs from the general image of shooting, in addition three
The topography from shooting of a picture element matrix;Matrix size is 92*92*6;
First convolutional layer extracts characteristics of image, convolution kernel for carrying out convolution operation with the picture element matrix of the input layer
5*5*6*20 is selected, biasing size is 20*1, show that characteristic pattern size is 88*88*20;
First pond layer, the characteristics of image picture element matrix for extracting to first convolutional layer carry out dimensionality reduction operation, mention
Notable feature information is taken, selection convolution kernel size is 2*2, and the characteristic pattern size after dimensionality reduction is 44*44*20;
Second convolutional layer, the characteristic pattern for obtaining to first pond layer further carry out convolution operation, extract
Characteristics of image, convolution kernel select 5*5*6*24, and biasing size is 24*1, show that characteristic pattern size is 40*40*24;
Second pond layer, the further dimensionality reduction of characteristic pattern for obtaining to second convolutional layer extract notable feature letter
Breath, selection convolution kernel size are 2*2, and the characteristic pattern size after dimensionality reduction is 20*20*24;
It is special to extract image for the third convolutional layer, the further convolution operation of characteristic pattern for obtaining to second pond layer
Sign, convolution kernel select 5*5*6*20, and biasing size is 24*1, show that characteristic pattern size is 16*16*24;
Third pond layer, the further dimensionality reduction operation of characteristic pattern for obtaining to the third convolutional layer, is extracted significant special
Reference breath, selection convolution kernel size are 2*2, and characteristic pattern size is 4*4*24 after dimensionality reduction;
The full articulamentum, the characteristics of image for obtaining to third pond layer are integrated, and weight matrix size is p*
192, it is biased to p*1;
The output layer, for exporting identification classification results, output matrix size is p*1, which represents right in input sample
The other probability of p kind herbs Class answered, the classification where being maximized are classification results.
3. the method as described in claim 1, characterized in that for two different training images of the equal shooting angle of every kind of medicinal material
Specific steps are as follows:
Each medicinal material is shot using drawer box, the drawer box, including square box ontology, the rectangular box
Sub- ontology includes top surface, bottom surface, left side, right side, leading flank and trailing flank, is all provided on the inside of the left side and right side
There is sliding rail, the sliding rail is separately positioned on the middle position of the vertical direction of left side and right side;The square box ontology
Bosom position is equipped with drawer, is separately mounted on the sliding rail on the inside of left side and right side at left and right sides of the drawer;
Square box ontology is divided into two spaces up and down by the drawer, and the square box top surface is equipped with rectangular slot, the square
Shape is slotted for placing mobile phone;The mobile phone is connect with background server;
Before shooting, medicinal material is put into drawer, after closing drawer, starts the photographing mode of mobile phone, the medicinal material in drawer is shot
First training image;
Then, drawer is extracted out from square box ontology, medicinal material is placed on square box bottom surface, start taking pictures for mobile phone
The medicinal material of mode, square shaped box bottom surface shoots second training image;
First training image is the topography for showing medicinal material regional area, and second training image is display medicinal material entirety area
The general image in domain;Double scale images refer to that the image source data of input is two kinds of entirety and part under equal resolution
The image in the visual field.
4. method as claimed in claim 3, characterized in that
The drawer box, when drawer is closed, entire drawer is sealing, due to there is about one push-and-pull outside drawer
The door of formula, so entire cassette interior is still sealing after drawer is taken away, as medicinal material establishes a camera bellows environment, to keep away
Exempt from ambient light and enter drawer cassette interior, and then interferes the shooting of medicinal material;
The drawer box, top surface are equipped with the lamp bead of S-shaped arrangement, and the lamp bead is powered by battery, the light that the lamp bead issues
For collimated white light, to guarantee that drawer box internal brightness is consistent;
The drawer box uses the resin material of white, and the opaque paster of one layer of black is posted outside drawer box, is avoided
Light is irradiated to drawer cassette interior, generates reflective influence to taking pictures;
The mobile phone is placed on rectangular slot.
5. the method as described in claim 1, characterized in that carry out pretreated tool to two training images of every kind of medicinal material
Body step are as follows:
For the quantity of spread training image, random process is carried out to the image of shooting, and to the image obtained after random process
With the training image before processing together as training image;The random process, comprising: Random-Rotation or overturning;Arbitrary color tone
Transformation, Random-fuzzy or random light shift field.
6. the method as described in claim 1, characterized in that make two training images of pretreated every kind of medicinal material side by side
It for one group of input value, is input in convolutional neural networks, using classification of medicinal material label as the output valve of convolutional neural networks, to volume
Product neural network is trained, and obtains the specific steps of trained convolutional neural networks are as follows:
One group of training image of input, including two different training images of shooting angle, every training image includes three defeated
Enter matrix;It includes six input matrixes that two training images, which have altogether,;
The one group of image and convolution kernel of input carry out convolution operation, and 1 input matrix corresponds to m convolution kernel by convolution operation i.e.
It show that m characteristic pattern, 6 input matrixes correspond to 6*m convolution kernel, obtains 6*m characteristic pattern by convolution operation, it will be described
In 6*m characteristic pattern, corresponding every 6 characteristic patterns carry out Fusion Features, that is, add up to 1 characteristic pattern, that is, finally obtain m spy
Sign figure;
Pond layer takes average operation to gained characteristic pattern, m picture pixels matrix after final output Fusion Features, and matrix passes through
The Classification and Identification of full articulamentum, exports the identification probability of p kind medicinal material, according to output result compared with actual result, obtains defeated
Layer error out is successively fed back, each layer parameter of regulating networks with gradient descent method by error back propagation;
Repetitive exercise several times is carried out, parameter is adjusted, when the loss function value of convolutional neural networks is less than given threshold
Or the number of iterations reach setting number when, deconditioning obtains trained convolutional neural networks.
7. the method as described in claim 1, characterized in that extract two to mapping using trained convolutional neural networks
The specific steps of result after the multi-features of picture are as follows:
S11:1 input matrix of step corresponds to m convolution kernel and obtains m characteristic pattern, and one group of testing image is that 6 input matrixes are
Corresponding 6*m convolution kernel obtains 6*m characteristic pattern, and the corresponding every 6 progress Fusion Features of characteristic pattern are added up to 1 feature
Figure, finally obtains m characteristic pattern;Pond layer takes average operation to gained characteristic pattern, m picture after final output Fusion Features
Picture element matrix;
Step S12: m picture pixels matrix after Fusion Features is added into biasing, and then each element of gained matrix is defeated
Enter into sigmoid activation primitive, extracts the profile and profile each section position feature of image;Subsequently into pond layer into
The dimensionality reduction of row picture element matrix, prominent main feature;
Step S13: by the feature extraction and selection of all convolutional layers and pond layer, the fusion of feature is realized;Finally by gained
Eigenmatrix pulls into a column vector in sequence.
8. the method as described in claim 1, characterized in that obtain the corresponding medicine of the testing image according to fused result
The specific steps of the recognition result of material are as follows:
Step S21: pulling into a column vector for gained eigenmatrix, is sent into full articulamentum, and the column vector is 192 dimensions;
Step S22: it by column vector multiplication corresponding with each row element of weight matrix, then is added, show that matrix size is
P*1, weight matrix size are p*192;
Step S23: it by gained matrix plus biasing, and then is sent into sigmoid activation primitive;
Step S24: by activation primitive output p as a result, representing the other probability of p kind herbs Class, the classification where being maximized
For final result.
9. the Chinese traditional medicinal materials recognition system based on double scale multi-features, characterized in that include: mobile phone, include depositing in mobile phone
Reservoir and processor, computer instruction is stored in the memory, and the processor is stored in memory for handling
Computer instruction is completed at the same time step described in any one of the claims 1-8 method.
10. a kind of computer readable storage medium, characterized in that operation has computer instruction, the computer instruction quilt thereon
When processor is run, step described in any one of the claims 1-8 method is completed.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201811525966.1A CN109615010B (en) | 2018-12-13 | 2018-12-13 | Traditional Chinese medicine material identification method and system based on double-scale convolutional neural network |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201811525966.1A CN109615010B (en) | 2018-12-13 | 2018-12-13 | Traditional Chinese medicine material identification method and system based on double-scale convolutional neural network |
Publications (2)
Publication Number | Publication Date |
---|---|
CN109615010A true CN109615010A (en) | 2019-04-12 |
CN109615010B CN109615010B (en) | 2020-11-10 |
Family
ID=66008249
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201811525966.1A Active CN109615010B (en) | 2018-12-13 | 2018-12-13 | Traditional Chinese medicine material identification method and system based on double-scale convolutional neural network |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN109615010B (en) |
Cited By (9)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110110728A (en) * | 2019-05-16 | 2019-08-09 | 湖南睿图智能科技有限公司 | A kind of Chinese traditional medicinal materials recognition method |
CN110596166A (en) * | 2019-09-11 | 2019-12-20 | 西京学院 | Method for identifying type and content of oil-gas reservoir space |
CN111337496A (en) * | 2020-04-13 | 2020-06-26 | 黑龙江北草堂中药材有限责任公司 | Chinese herbal medicine picking device and picking method |
CN111353432A (en) * | 2020-02-28 | 2020-06-30 | 安徽华润金蟾药业股份有限公司 | Rapid honeysuckle medicinal material cleaning method and system based on convolutional neural network |
CN112164076A (en) * | 2020-09-24 | 2021-01-01 | 济南大学 | Hardened cement water cement ratio prediction method and system based on cement microstructure image |
CN114789870A (en) * | 2022-05-20 | 2022-07-26 | 深圳市信成医疗科技有限公司 | Innovative modular drug storage management implementation mode |
CN114998639A (en) * | 2022-04-19 | 2022-09-02 | 安徽农业大学 | Chinese medicinal material class identification method based on deep learning |
WO2024001467A1 (en) * | 2022-07-01 | 2024-01-04 | 湖南大学 | Fatigue crack propagation rate test method and device based on deep learning |
CN117953354A (en) * | 2024-03-22 | 2024-04-30 | 深圳禾思众成科技有限公司 | Oversized pixel image processing method and system for computing unified equipment architecture |
Citations (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US6633682B1 (en) * | 1999-06-21 | 2003-10-14 | New York University | Progressive fractal rendering |
CN201826835U (en) * | 2009-07-13 | 2011-05-11 | 新疆鑫诚信石油技术服务有限公司 | Rock debris image acquisition and analysis device |
CN107392224A (en) * | 2017-06-12 | 2017-11-24 | 天津科技大学 | A kind of crop disease recognizer based on triple channel convolutional neural networks |
CN107958257A (en) * | 2017-10-11 | 2018-04-24 | 华南理工大学 | A kind of Chinese traditional medicinal materials recognition method based on deep neural network |
CN108256568A (en) * | 2018-01-12 | 2018-07-06 | 宁夏智启连山科技有限公司 | A kind of plant species identification method and device |
-
2018
- 2018-12-13 CN CN201811525966.1A patent/CN109615010B/en active Active
Patent Citations (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US6633682B1 (en) * | 1999-06-21 | 2003-10-14 | New York University | Progressive fractal rendering |
CN201826835U (en) * | 2009-07-13 | 2011-05-11 | 新疆鑫诚信石油技术服务有限公司 | Rock debris image acquisition and analysis device |
CN107392224A (en) * | 2017-06-12 | 2017-11-24 | 天津科技大学 | A kind of crop disease recognizer based on triple channel convolutional neural networks |
CN107958257A (en) * | 2017-10-11 | 2018-04-24 | 华南理工大学 | A kind of Chinese traditional medicinal materials recognition method based on deep neural network |
CN108256568A (en) * | 2018-01-12 | 2018-07-06 | 宁夏智启连山科技有限公司 | A kind of plant species identification method and device |
Cited By (11)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110110728A (en) * | 2019-05-16 | 2019-08-09 | 湖南睿图智能科技有限公司 | A kind of Chinese traditional medicinal materials recognition method |
CN110596166A (en) * | 2019-09-11 | 2019-12-20 | 西京学院 | Method for identifying type and content of oil-gas reservoir space |
CN111353432A (en) * | 2020-02-28 | 2020-06-30 | 安徽华润金蟾药业股份有限公司 | Rapid honeysuckle medicinal material cleaning method and system based on convolutional neural network |
CN111353432B (en) * | 2020-02-28 | 2023-08-01 | 安徽华润金蟾药业股份有限公司 | Rapid clean selection method and system for honeysuckle medicinal materials based on convolutional neural network |
CN111337496A (en) * | 2020-04-13 | 2020-06-26 | 黑龙江北草堂中药材有限责任公司 | Chinese herbal medicine picking device and picking method |
CN112164076A (en) * | 2020-09-24 | 2021-01-01 | 济南大学 | Hardened cement water cement ratio prediction method and system based on cement microstructure image |
CN114998639A (en) * | 2022-04-19 | 2022-09-02 | 安徽农业大学 | Chinese medicinal material class identification method based on deep learning |
CN114998639B (en) * | 2022-04-19 | 2024-04-26 | 安徽农业大学 | Deep learning-based traditional Chinese medicine category identification method |
CN114789870A (en) * | 2022-05-20 | 2022-07-26 | 深圳市信成医疗科技有限公司 | Innovative modular drug storage management implementation mode |
WO2024001467A1 (en) * | 2022-07-01 | 2024-01-04 | 湖南大学 | Fatigue crack propagation rate test method and device based on deep learning |
CN117953354A (en) * | 2024-03-22 | 2024-04-30 | 深圳禾思众成科技有限公司 | Oversized pixel image processing method and system for computing unified equipment architecture |
Also Published As
Publication number | Publication date |
---|---|
CN109615010B (en) | 2020-11-10 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN109615010A (en) | Chinese traditional medicinal materials recognition method and system based on double scale convolutional neural networks | |
CN109615574A (en) | Chinese medicine recognition methods and system based on GPU and double scale image feature comparisons | |
CN106204779B (en) | Check class attendance method based on plurality of human faces data collection strategy and deep learning | |
CN106469302B (en) | A kind of face skin quality detection method based on artificial neural network | |
CN108229490A (en) | Critical point detection method, neural network training method, device and electronic equipment | |
CN109508650A (en) | A kind of wood recognition method based on transfer learning | |
CN107463919A (en) | A kind of method that human facial expression recognition is carried out based on depth 3D convolutional neural networks | |
CN108806792A (en) | Deep learning facial diagnosis system | |
CN107610123A (en) | A kind of image aesthetic quality evaluation method based on depth convolutional neural networks | |
CN106372581A (en) | Method for constructing and training human face identification feature extraction network | |
CN109815893A (en) | The normalized method in colorized face images illumination domain of confrontation network is generated based on circulation | |
CN107945118A (en) | A kind of facial image restorative procedure based on production confrontation network | |
CN110309880A (en) | A kind of 5 days and 9 days hatching egg embryo's image classification methods based on attention mechanism CNN | |
CN109086723A (en) | A kind of method, apparatus and equipment of the Face datection based on transfer learning | |
CN109886153A (en) | A kind of real-time face detection method based on depth convolutional neural networks | |
CN109190643A (en) | Based on the recognition methods of convolutional neural networks Chinese medicine and electronic equipment | |
CN104298974A (en) | Human body behavior recognition method based on depth video sequence | |
CN110348375A (en) | A kind of finger vena region of interest area detecting method neural network based | |
CN104298973A (en) | Face image rotation method based on autoencoder | |
CN107633229A (en) | Method for detecting human face and device based on convolutional neural networks | |
CN109871845A (en) | Certificate image extracting method and terminal device | |
CN110232326A (en) | A kind of D object recognition method, device and storage medium | |
CN108664843A (en) | Live subject recognition methods, equipment and computer readable storage medium | |
CN110472495A (en) | A kind of deep learning face identification method based on graphical inference global characteristics | |
Fu et al. | Computerized tongue coating nature diagnosis using convolutional neural network |
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 | ||
GR01 | Patent grant | ||
GR01 | Patent grant |