CN110084215A - A kind of pedestrian of the twin network model of binaryzation triple recognition methods and system again - Google Patents
A kind of pedestrian of the twin network model of binaryzation triple recognition methods and system again Download PDFInfo
- Publication number
- CN110084215A CN110084215A CN201910369005.4A CN201910369005A CN110084215A CN 110084215 A CN110084215 A CN 110084215A CN 201910369005 A CN201910369005 A CN 201910369005A CN 110084215 A CN110084215 A CN 110084215A
- Authority
- CN
- China
- Prior art keywords
- sample
- binaryzation
- layer
- weight
- pedestrian
- 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
- 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
- G06V40/00—Recognition of biometric, human-related or animal-related patterns in image or video data
- G06V40/10—Human or animal bodies, e.g. vehicle occupants or pedestrians; Body parts, e.g. hands
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- General Physics & Mathematics (AREA)
- General Health & Medical Sciences (AREA)
- General Engineering & Computer Science (AREA)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- Data Mining & Analysis (AREA)
- Evolutionary Computation (AREA)
- Artificial Intelligence (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- Biomedical Technology (AREA)
- Life Sciences & Earth Sciences (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Health & Medical Sciences (AREA)
- Human Computer Interaction (AREA)
- Multimedia (AREA)
- Image Analysis (AREA)
Abstract
The pedestrian that the present invention discloses a kind of twin network model of binaryzation triple identifies again and method, include: three road convolutional neural networks, input positive and negative samples and detection sample image respectively, extract characteristics of image, and every road convolutional neural networks include convolutional layer, pond floor and full articulamentum；Convolutional layer, the pond floor of every road convolutional neural networks are identical with full articulamentum and weighting parameter is shared, and carry out binaryzation to weighting parameter and to the activation primitive value between convolutional layer and pond layer；It Softmax layers, is connect with each full articulamentum, the feature that convolutional neural networks export is classified and normalized；Triplet loss verifies function module, connect with Softmax layers, receives the sample characteristics of normalization classification layer output, carries out the Similarity measures of sample pair.The present invention obtains the stronger deep learning model of distinguishing ability, better solves the problem that similar and identical pedestrian's picture causes content deltas big because of illumination, scene changes, the diversified reason of human body attitude again on different pedestrian's image contents.
Description
Technical field
The present invention relates to computer visual image and field of video processing, in particular to a kind of twin net of binaryzation triple
The pedestrian of network model recognition methods and system again.
Background technique
Pedestrian identifies again can be generally considered as an image retrieval problem, i.e., matches pedestrian from different cameras.
The pedestrian image of a given inquiry, pedestrian identify again be intended to find out from the pedestrian image library under non-overlap camera angles and
The pedestrian is the image of the same pedestrian.
Due to video camera capture pedestrian image by illumination, posture, visual angle, image resolution ratio, camera setting, block and
Background is equal in a jumble to be influenced, it may appear that the same pedestrian image, which changes a lot, wears same clothes without same pedestrian or have similar
The phenomenon that same person is identified as when appearance instead.
Therefore, pedestrian identifies to be still a challenging task in computer vision again.Since 2012, with volume
Product neural network (CNN) is that the deep learning model of representative achieves immense success in computer vision field.Meanwhile it also driving
Pedestrian again identifies the research in field.Compared to pedestrian's recognition methods again of traditional-handwork design pedestrian's feature, the row based on CNN
Recognition methods can more efficiently overcome the complicated variation of pedestrian to people again, achieve higher performance.But it is current based on
The pedestrian of CNN again recognition methods still not can be well solved on different pedestrian's image contents similar and identical pedestrian's picture again because
The problem for causing content deltas big for illumination, scene changes, the diversified reason of human body attitude.
Summary of the invention
The purpose of the present invention is to provide a kind of pedestrian of twin network model of binaryzation triple again recognition methods and it is
System devises weight shared associated losses functional expression triple network model on the basis of general Siamese model, by
In detection pedestrian image by illumination, posture, visual angle, image resolution ratio, camera setting, block and background is mixed and disorderly etc. influences,
It will appear the same pedestrian image and change a lot and wear same clothes without same pedestrian or identified instead when having similar appearance
The phenomenon that for the same person, the network structure combining classification loss function Softmax loss proposed by the present invention and verifying function
Training may finally obtain the stronger depth of distinguishing ability to construct the network model that pedestrian identifies again to Triplet loss simultaneously
Learning model can better solve similar and identical pedestrian's picture on different pedestrian's image contents and become again because of illumination, scene
Change, the problem that the diversified reason of human body attitude causes content deltas big.
In order to achieve the above object, the invention is realized by the following technical scheme:
A kind of pedestrian's weight identifying system of the twin network model of binaryzation triple includes:
Three road convolutional neural networks input positive sample, negative sample and detection sample image respectively, extract corresponding image
Feature；Each convolutional neural networks include: input layer, extract the feature of sample image pair；Convolutional layer connects with the input layer
It connects, for extracting the characteristics of image of described image sample；Pond layer, connect with the convolutional layer, obtains spy by dimension-reduction treatment
Determine characteristic area, and be integrated into the feature vector of specific dimension values, is sent to full articulamentum；Wherein, the convolution mind on every road
Convolutional layer, pond layer through network be identical with full articulamentum and weighting parameter is shared, and to weighting parameter carry out binaryzation and
Binaryzation is carried out to the activation primitive value in every road convolutional neural networks between convolutional layer and pond floor；
Normalization classification layer, connect with each full articulamentum of three convolutional neural networks, by the convolution mind
The feature exported through network is classified and is normalized；
Error loss verifying function module, connect with normalization classification layer, and it is defeated to receive the normalization classification layer
Sample characteristics out carry out the Similarity measures of sample pair.
Preferably, the normalization classification layer is Softmax layers；
The formula of described Softmax layers of Softmax Loss function are as follows:
Wherein, L is loss；Y refers to vector；Sj is j-th of value of softmax layers of output vector S, and expression is current
Sample belongs to the probability of j-th of classification.
Preferably, the error function module is that Triplet loss verifies function；
The method of the Triplet loss verifying function are as follows:
Select image and negative sample of the picture as detection sample, setting positive sample and detection sample for the same person
For the image of the different people of detection sample, building obtains the triple being made of detection sample, positive sample and negative sample；
By the twin network model of the triple, keeps the detection sample closer at a distance from positive sample and make institute
It is farther at a distance from the negative sample to state detection sample；
When the positive sample example distance and detection the distance between sample closely to set distance when, then the detection sample
Picture and positive sample picture be same a group traveling together, when the negative sample distance and detection the distance between sample are as far as set distance
When, then the picture of the detection sample and the picture of negative sample are not go together.
Preferably, in the binarization, by the activation primitive value binaryzation between weight and convolutional layer and pond layer
It is 1 or -1, using the Deterministic Methods based on sign function Sign, is shown below:
Wherein, for activation primitive value, the value of binaryzation directly is obtained using the binaryzation function of formula (2)；For weight,
Part by will exceed [- 1,1] when updating weighting parameter is cut, and to be maintained between [- 1,1], reuses formula (2)
Binaryzation function obtain the value of binaryzation.
Preferably, pedestrian's weight identifying system of the twin network model of binaryzation triple further includes following mistake
Journey: after obtaining the activation primitive value of any layer, when using weight parameter, by weight binaryzation, then with preceding layer neuron
Activation primitive value after node binaryzation is multiplied, and so on carry out the propagation and update of weight parameter, then carry out batch standardization
The input for making each layer of neural network in convolutional neural networks training process is set to keep same distribution.
Preferably, the parameter transformation in the binarization of the weight and update step are as follows:
S1, start binaryzation；
S2, judge whether iteration is completed, if so, terminating, if it is not, then continuing to execute step S3；
S3, binaryzation is carried out to every layer of weight and calculates loss；
S4, loss is calculated to the derivative of binaryzation weight；
S5, floating-point weight is updated, and jumps to the step S2, until iteration terminates.
The present invention also provides a kind of pedestrians based on the twin network model of binaryzation triple as described above to know again
Other system,
Extract image pair characteristics of image, described image to include positive sample, negative sample and detection sample image；
Three road convolutional neural networks of the twin network model of triple extract positive sample, negative sample and detection sample graph respectively
The feature of picture is trained, the characteristic value exported, and according to the gap of this feature value and setting value, adjusts the weight of model
Matrix, the twin network model of triple after being trained；Wherein, convolutional layer, the Chi Hua of the convolutional neural networks on every road
Layer is identical with full articulamentum and weighting parameter is shared, and to weighting parameter progress binaryzation and in every road convolutional neural networks
Activation primitive value between convolutional layer and pond layer carries out binaryzation；
Classification normalization is carried out to characteristic value, obtains the mark sheet of image；
Will test sample it is closer at a distance from positive sample and make detect sample it is farther at a distance from negative sample；Work as positive sample
When the distance between example distance and detection sample closely arrive set distance, then the picture and positive sample picture for detecting sample are same
Pedestrian then detects the picture and negative sample of sample when the distance between negative sample distance and detection sample are as far as set distance
Picture be different pedestrians.
Compared with prior art, the invention has the benefit that
(1) the twin network model training of triple of the invention produces more compared with the twin network of previous two branch
The triplets of more picture numbers, thus while training, can effectively alleviate over-fitting；The present invention passes through setting three networks point
The training pattern that identifies again of parameter sharing building pedestrian of branch can better solve similar on different pedestrian's image contents and phase
With pedestrian's picture again because of the problem that the reasons such as illumination, scene changes cause content deltas big, rolled up with traditional depth end to end
Product neural network model is compared, and the high-precision classification effect of disaggregated model has not only been remained, but also adequately achieves verifying model Gao Zhun
The advantages of judgement to identify whether as same a group traveling together of exactness, improves the comprehensive performance that pedestrian identifies again, has to learn one
Very strong discerning CNN feature and similarity measurement；
(2) existing neural network generally uses Floating-point Computation, needs biggish memory space and calculation amount, serious to hinder
Application in mobile terminal；Binaryzation neural network proposed by the present invention have high model compression ratio and fast calculating speed it is potential excellent
Gesture, after carrying out binary conversion treatment to weight W and activation primitive value, it is ensured that pedestrian retrieval loss in accuracy is lesser to be discharged simultaneously
Memory is to reduce computer loss.
Detailed description of the invention
Fig. 1 is the whole twin network structure of associated losses functional expression triple of the invention；
Fig. 2 is Softmax principle of classification figure of the invention；
Fig. 3 is Triplet loss distance verifying schematic diagram of the invention；
Fig. 4 is the schematic network structure of associated losses function of the invention；
Fig. 5 is that binaryzation network parameter of the invention is propagated and updates flow chart.
Specific embodiment
By reading detailed description of non-limiting embodiments made by-Fig. 5 referring to Fig.1, feature of the invention,
Objects and advantages will become more apparent upon.Referring to Fig. 1-Fig. 5 for showing the embodiment of the present invention, this hair hereafter will be described in greater detail
It is bright.However, the present invention can be realized by many different forms, and it should not be construed as the limit by the embodiment herein proposed
System.
It is as shown in Figure 1 the overall structure figure of the twin network of binaryzation triple of the invention.Binaryzation three of the invention
Pedestrian's weight identifying system of the twin network model of tuple includes the very strong CNN model of three judgement index.In Fig. 1, input picture sequence
Column are very multiple series of images pair, for training and testing；Each image is to mainly by positive sample (negative), negative sample
(negative) it is formed with detection sample (anchor).The image pair that size is set when given n, selects suitable network
Network structure of the model as CNN model, three CNN models in the twin network of triple receive positive sample, negative sample respectively
With detection sample image information, and then extract image feature.
Each CNN (convolutional neural networks) model mainly includes input layer, convolutional layer, activation primitive, pond layer and Quan Lian
Connect layer.Wherein, input layer receives each image of input, and further extracts characteristics of image using the convolutional layer in CNN model,
Pond layer carries out dimension-reduction treatment and obtains special characteristic region, then continues to for being sent into CNN model in special characteristic region, in pond
The maximum that is followed by of layer be averaged in the layer of pond be by special characteristic Regional Integration specific dimension values feature vector, and output is connected entirely
Connect layer.
As shown in Fig. 1 and Fig. 4 combination, each full articulamentum of CNN model is connected with Softmax layers, the Softmax layers of progress
Normalization, which classifies and exports feature, verifies function to Triplet loss, and the two, which combines, to be trained, to can be obtained more
Good detection accuracy.
CNN model of the invention can be used the networks such as network VGGnet-16, Resnet-50 and CaffeNet and be instructed
Practice, and can in many public databases (such as Market-1501 data set, MARS data set, CUHK03 data set) into
Row experiment.Wherein, the present invention realizes proposed method using deep learning tensor library PyTorch.In order to verify this method
Validity, the present invention selects VGGnet-16 network to test on Market-1501 data set, tests hardware used
Configure as follows, GPU:Nvidia GeForce GTX 1080Ti (video memory: 12GB), memory: 128GB, CPU:Intel Core eight
Core i7 processor (dominant frequency 3.60GHz).The experiment further demonstrates method of the invention and is guaranteeing that detection accuracy loss is smaller
In the case where, original 1/32 times is reduced on memory consumption, this greatly releases calculator memory, reduces computer damage
Consumption.
It is illustrated in figure 2 Softmax principle of classification figure of the invention.W41, w42, w43, w51, w52, w53 in Fig. 2,
W61, w62, w63 are parameters, 1,2,3,4,5,6 in Fig. 2 be respectively each node, for example, if desired find out w41,
The local derviation of w42, w43 then need to ask local derviation to pass to node 4 Softmax Loss function, then chain rule are recycled to continue
The local derviation of derivation, the parameters such as w51 ... w63 can similarly be found out.
Wherein, Softmax Loss function (i.e. the loss function of Softmax) is often being classified in convolutional neural networks more
Scene in using extensively, formula are as follows:
Wherein, L is loss；yjThere is a summation symbol in front, and the range of j is also 1 to classification number T, therefore y is a 1*T
Vector, T of the inside value, and only 1 value is 1, other T-1 value is all 0；Sj is softmax layers of output vector S
J-th of value, expression is probability that current sample belongs to j-th of classification.
It is illustrated in figure 3 Triplet loss distance verifying schematic diagram of the invention.The present invention passes through optimization sample instantiation
(anchor, also referred to as anchor example) and positive (Positive) exemplary distance be less than sample instantiation and negative (Negative) it is exemplary away from
From come the Similarity measures of realizing sample pair.Specifically, select a picture as anchor (detection sample), Positive
(positive sample) is with the anchor same person, and Negative (negative sample) is with the different people of anchor.Of the invention
The twin network of network model of triple by study so that anchor is closer at a distance from positive, and make anchor with
The distance of negative is far.If positive example distance and sample instantiation are apart from close, then it is assumed that the samples pictures and positive sample
This picture is same a group traveling together；If negative example distance and sample instantiation are apart from far, then it is assumed that the samples pictures and negative sample figure
Piece is different pedestrians, judges whether pedestrian is same a group traveling together with this distance.
Wherein, Triplet loss principle is as follows:
XanchorIndicate anchor sample, XPositiveIndicate similar sample, XNegativeIndicate that similar sample, threshold indicate
Specific threshold, the inequality can be expressed as form:
The inequality substantially defines the distance between similar sample and foreign peoples's sample relationship, it may be assumed that all similar samples
The distance between+threshold value threshold is less than the distance between foreign peoples's sample.
When distance relation is unsatisfactory for above-mentioned inequality, back-propagation method can be passed through by solving following objective functions
Entire whole network is adjusted, objective function is as follows:
As long as the value of formula is greater than 0 in bracket, error is just calculated, X can be calculated separately out using the objective functionanchor,
XPositive, XNegativeGradient direction, and according to back-propagation method adjust front convolutional neural networks, convolutional Neural net
Network continues training, accurate adjustment, and circulation carries out, until meeting the above-mentioned inequality of setting, judges whether pedestrian is same with this
Pedestrian.
In the present embodiment, the convolutional layer of every road CNN model, the dimension of pond floor and full articulamentum, structure and weight setting
All identical and weighting parameter is shared, it is possible to reduce the quantity of parameter, and have activation primitive value between every road convolutional layer and pond floor
Carrying out imparametrization processing keeps classifying quality more preferable.
Wherein, the connection type of neural network between layers is whole neuron phases of each neuron with upper one layer
Even, the weight of these connecting lines is independently of other neurons, it is assumed that upper one layer is m neuron, current layer is n mind
Through member, then shared m × n connection, also just has m × n weight, weight matrix is exactly m × n shape.Generally use weight matrix W
It indicates, every a line is the value for the weight that a neuron is connected with upper one layer of all neuron.When, vacation shared without using weight
If characteristic pattern is made of the characteristic pattern of 10 32*32*1, there are 1024 neurons on each characteristic pattern, each neuron corresponds to defeated
Enter the region of one piece of 5*5*3 on image, so there are 75 connections in this block region of a neuron and input picture, i.e., 75
Weighting parameter then shares 75*1024*10=768000 weighting parameter, this is very unfavorable for the adjustment and transmitting of parameter, because
This convolutional neural networks introduces weight and shares principle, and corresponding 75 weighting parameters of each neuron are each on a characteristic pattern
Neuron is shared, and 75 weighting parameters use corresponding identical weight in the shared each neuron just referred in fact here,
Weight just only needs 75*10=750 weighting parameter when shared, and the threshold value of each characteristic pattern is also shared, that is, needs 10 threshold values,
750+10=760 parameter is then needed in total.
So the weight due to convolutional layer, pond layer and full articulamentum is shared, then the binaryzation behaviour of weight is directly carried out
Make, and the activation primitive value in the CNN model of the road Ye Kegeimei between convolutional layer and pond floor carries out binarization operation.
The binarization operation principle is as follows:
Weight W and activation primitive value two-value are turned to 1 or -1, the parameter of CNN model is made to occupy smaller memory space.Its
In, it is -1 less than 0 that using certainty (deterministic) method based on sign function Sign, that is, being greater than 0, which is just+1,
It is shown below:
Wherein, for activation primitive value, the value of binaryzation directly is obtained using the binaryzation function of decision formula (2).In addition, right
In weight, it is also necessary in undated parameter beyond [- 1,1] part to cut, i.e., holding weight parameter always [- 1,
1] it between, recycles and obtains the value of binaryzation using the binaryzation function of decision formula (2).
From the foregoing, it will be observed that after obtaining the activation primitive value of a certain layer, when using weight parameter, need by weight parameter into
Row binaryzation, first first by the weight W binaryzation of neural network, then with the activation letter after preceding layer neuron node binaryzation
Numerical value is multiplied, and so on carry out the propagation and update of weight parameter, then carry out BatchNormalization (batch standardization)
The input for making each layer of neural network in deep neural network training process is set to keep same distribution.
The binaryzation network parameter for being illustrated in figure 5 weight parameter is propagated and updates flow chart comprising the steps of:
S1, start binaryzation；
S2, judge whether iteration is completed, if so, terminating, calculate loss simultaneously if it is not, thening follow the steps S3 and continuing binaryzation
Undated parameter readjustment is trained；Wherein, judge that iteration completion is equivalent to judge whether network model trains and finish, it can also
It is interpreted as whether the network model training precision has reached requirement；
S3, binaryzation is carried out to every layer of weight and calculates loss (loss)；
S4, loss is calculated to the derivative of binaryzation weight；
S5, floating-point weight is updated, and the S2 that gos to step, until the process terminates.
The present invention by Softmax layers of Classification Loss function Softmax loss and verifying function Triplet loss this two
Person is combined together carry out network training, and the network model progress pedestrian that can train efficiently classification and high-precision verifying knows again
Not.Wherein, Softmax layers with Triplet loss loss function by Python program language combine realization instruct together
Practice, while classification can be further realized in conjunction with the validity of verifying function.The present invention and traditional depth end to end are rolled up
Product neural network model is compared, and the high-precision classification effect of disaggregated model has not only been remained, but also adequately achieves verifying model Gao Zhun
The advantages of judgement to identify whether as same a group traveling together of exactness, improves the comprehensive performance that pedestrian identifies again.The present invention passes through two-value
Change operation binaryzation weight, weight W and hidden layer activation value two-value are turned to 1 or -1, the parameter of model is made to occupy smaller deposit
Space is stored up, is furthermore achieved in the case where guaranteeing the lesser situation of loss of significance, computer loss is reduced.
It is discussed in detail although the contents of the present invention have passed through above preferred embodiment, but it should be appreciated that above-mentioned
Description is not considered as limitation of the present invention.After those skilled in the art have read above content, for of the invention
A variety of modifications and substitutions all will be apparent.Therefore, protection scope of the present invention should be limited to the appended claims.
Claims (7)
1. a kind of twin network model of binaryzation triple pedestrian weight identifying system, characterized by comprising:
Three road convolutional neural networks input positive sample, negative sample and detection sample image respectively, it is special to extract corresponding image
Sign；Each convolutional neural networks include: input layer, extract the feature of sample image pair；Convolutional layer connects with the input layer
It connects, for extracting the characteristics of image of described image sample；Pond layer, connect with the convolutional layer, obtains spy by dimension-reduction treatment
Determine characteristic area, and be integrated into the feature vector of specific dimension values, is sent to full articulamentum；Wherein, the convolution mind on every road
Convolutional layer, pond layer through network be identical with full articulamentum and weighting parameter is shared, and to weighting parameter carry out binaryzation and
Binaryzation is carried out to the activation primitive value in every road convolutional neural networks between convolutional layer and pond floor；
Normalization classification layer, connect, by the convolutional Neural net with each full articulamentum of three road convolutional neural networks
The feature of network output is classified and is normalized；
Error loss verifying function module, connect with normalization classification layer, receives the normalization classification layer output
Sample characteristics carry out the Similarity measures of sample pair.
2. pedestrian's weight identifying system of the twin network model of binaryzation triple as described in claim 1, which is characterized in that
The normalization classification layer is Softmax layers；
The formula of described Softmax layers of Softmax Loss function are as follows:
Wherein, L is loss；Y refers to vector；Sj is j-th of value of softmax layers of output vector S, and expression is current sample
Belong to the probability of j-th of classification.
3. pedestrian's weight identifying system of the twin network model of binaryzation triple as described in claim 1, which is characterized in that
The error function module is that Triplet loss verifies function；
The method of the Triplet loss verifying function are as follows:
Select a picture as detection sample, setting positive sample is the image of the same person with detection sample and negative sample is inspection
The image of this different people of test sample, building obtain the triple being made of detection sample, positive sample and negative sample；
By the twin network model of the triple, keeps the detection sample closer at a distance from positive sample and make the inspection
Test sample sheet is farther at a distance from the negative sample；
When the positive sample example distance and detection the distance between sample closely to set distance when, then the figure of the detection sample
Piece and positive sample picture are same a group traveling together, when the distance between the negative sample distance and detection sample are as far as set distance,
Then the picture of the detection sample is different pedestrians from the picture of negative sample.
4. pedestrian's weight identifying system of the twin network model of binaryzation triple as described in claim 1, which is characterized in that
In the binarization, the activation primitive value two-value between weight and convolutional layer and pond layer is turned to 1 or -1, is used
Based on the Deterministic Methods of sign function Sign, it is shown below:
Wherein, for activation primitive value, the value of binaryzation directly is obtained using the binaryzation function of formula (2)；
For weight, the part by will exceed [- 1,1] when updating weighting parameter is cut, be maintained at [- 1,1] it
Between, the binaryzation function for reusing formula (2) obtains the value of binaryzation.
5. pedestrian's weight identifying system of the twin network model of binaryzation triple as described in claim 1, which is characterized in that into
One step includes following procedure:
After obtaining the activation primitive value of any layer, when using weight parameter, by weight binaryzation, then with preceding layer neuron
Activation primitive value after node binaryzation is multiplied, and so on carry out the propagation and update of weight parameter, then carry out batch standardization
The input for making each layer of neural network in convolutional neural networks training process is set to keep same distribution.
6. pedestrian's weight identifying system of the twin network model of binaryzation triple as described in claim 4 or 5, feature exist
In,
Parameter transformation and update step in the binarization of the weight are as follows:
S1, start binaryzation；
S2, judge whether iteration is completed, if so, terminating, if it is not, then continuing to execute step S3；
S3, binaryzation is carried out to every layer of weight and calculates loss；
S4, loss is calculated to the derivative of binaryzation weight；
S5, floating-point weight is updated, and jumps to the step S2, until iteration terminates.
7. a kind of pedestrian based on the twin network model of binaryzation triple as claimed in any one of claims 1 to 6 identifies again
System, which is characterized in that
Extract image pair characteristics of image, described image to include positive sample, negative sample and detection sample image；
Three road convolutional neural networks of the twin network model of triple extract positive sample, negative sample and detection sample image respectively
Feature is trained, the characteristic value exported, and according to the gap of this feature value and setting value, adjusts the weight square of model
Battle array, the twin network model of triple after being trained；Wherein, the convolutional layer of the convolutional neural networks on every road, pond floor
And weighting parameter identical with full articulamentum is shared, and carries out binaryzation to weighting parameter and to rolling up in every road convolutional neural networks
Activation primitive value between lamination and pond layer carries out binaryzation；
Classification normalization is carried out to characteristic value, obtains the mark sheet of image；
Will test sample it is closer at a distance from positive sample and make detect sample it is farther at a distance from negative sample；When positive sample example
When the distance between distance and detection sample closely arrive set distance, then the picture and positive sample picture for detecting sample are same a line
People then detects the picture and negative sample of sample when negative sample distance and detection the distance between sample are as far as set distance
Picture is different pedestrians.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910369005.4A CN110084215A (en) | 2019-05-05 | 2019-05-05 | A kind of pedestrian of the twin network model of binaryzation triple recognition methods and system again |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910369005.4A CN110084215A (en) | 2019-05-05 | 2019-05-05 | A kind of pedestrian of the twin network model of binaryzation triple recognition methods and system again |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN110084215A true CN110084215A (en) | 2019-08-02 |
Family
ID=67418566
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201910369005.4A Pending CN110084215A (en) | 2019-05-05 | 2019-05-05 | A kind of pedestrian of the twin network model of binaryzation triple recognition methods and system again |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN110084215A (en) |
Cited By (23)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110503053A (en) * | 2019-08-27 | 2019-11-26 | 电子科技大学 | Human motion recognition method based on cyclic convolution neural network |
| CN110516745A (en) * | 2019-08-28 | 2019-11-29 | 北京达佳互联信息技术有限公司 | Training method, device and the electronic equipment of image recognition model |
| CN110532890A (en) * | 2019-08-05 | 2019-12-03 | 安徽大学 | A kind of twin convolutional neural networks pedestrian recognition methods again of distribution based on cloud, marginal end and equipment end |
| CN110580460A (en) * | 2019-08-28 | 2019-12-17 | 西北工业大学 | Pedestrian re-identification method based on combined identification and verification of pedestrian identity and attribute characteristics |
| CN110659585A (en) * | 2019-08-31 | 2020-01-07 | 电子科技大学 | Pedestrian detection method based on interactive attribute supervision |
| CN110728330A (en) * | 2019-10-23 | 2020-01-24 | 腾讯科技（深圳）有限公司 | Object identification method, device, equipment and storage medium based on artificial intelligence |
| CN110738241A (en) * | 2019-09-24 | 2020-01-31 | 中山大学 | binocular stereo vision matching method based on neural network and operation frame thereof |
| CN110837775A (en) * | 2019-09-30 | 2020-02-25 | 合肥合工安驰智能科技有限公司 | Underground locomotive pedestrian and distance detection method based on binarization network |
| CN110929099A (en) * | 2019-11-28 | 2020-03-27 | 杭州趣维科技有限公司 | Short video frame semantic extraction method and system based on multitask learning |
| CN111079585A (en) * | 2019-12-03 | 2020-04-28 | 浙江工商大学 | Image enhancement and pseudo-twin convolution neural network combined pedestrian re-identification method based on deep learning |
| CN111783624A (en) * | 2020-06-29 | 2020-10-16 | 厦门市美亚柏科信息股份有限公司 | Pedestrian re-identification method and device with translation invariance reserved and storage medium |
| CN111797700A (en) * | 2020-06-10 | 2020-10-20 | 南昌大学 | Vehicle re-identification method based on fine-grained discrimination network and second-order reordering |
| CN112434790A (en) * | 2020-11-10 | 2021-03-02 | 西安理工大学 | Self-interpretation method for convolutional neural network to judge partial black box problem |
| CN112508114A (en) * | 2020-12-14 | 2021-03-16 | 武汉纺织大学 | Intelligent clothing recommendation system and method |
| CN112949384A (en) * | 2021-01-23 | 2021-06-11 | 西北工业大学 | Remote sensing image scene classification method based on antagonistic feature extraction |
| CN113221530A (en) * | 2021-04-19 | 2021-08-06 | 杭州火石数智科技有限公司 | Text similarity matching method and device based on circle loss, computer equipment and storage medium |
| CN113256592A (en) * | 2021-06-07 | 2021-08-13 | 中国人民解放军总医院 | Training method, system and device of image feature extraction model |
| CN113362070A (en) * | 2021-06-03 | 2021-09-07 | 中国工商银行股份有限公司 | Method, apparatus, electronic device, and medium for identifying operating user |
| CN113361693A (en) * | 2021-06-30 | 2021-09-07 | 北京百度网讯科技有限公司 | Method and device for generating convolutional neural network and image recognition method and device |
| CN113435383A (en) * | 2021-07-07 | 2021-09-24 | 中国人民解放军国防科技大学 | Remote sensing airplane target classification method and device based on double triple pseudo-twin framework |
| CN113469253A (en) * | 2021-07-02 | 2021-10-01 | 河海大学 | Electricity stealing detection method based on triple twin network |
| CN113781518A (en) * | 2021-09-10 | 2021-12-10 | 商汤集团有限公司 | Neural network structure searching method and device, electronic device and storage medium |
| CN113963150A (en) * | 2021-11-16 | 2022-01-21 | 北京中电兴发科技有限公司 | Pedestrian re-identification method based on multi-scale twin cascade network |
Citations (14)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US5778152A (en) * | 1992-10-01 | 1998-07-07 | Sony Corporation | Training method for neural network |
| CN105354568A (en) * | 2015-08-24 | 2016-02-24 | 西安电子科技大学 | Convolutional neural network based vehicle logo identification method |
| CN105608450A (en) * | 2016-03-01 | 2016-05-25 | 天津中科智能识别产业技术研究院有限公司 | Heterogeneous face identification method based on deep convolutional neural network |
| CN105975931A (en) * | 2016-05-04 | 2016-09-28 | 浙江大学 | Convolutional neural network face recognition method based on multi-scale pooling |
| CN106778464A (en) * | 2016-11-09 | 2017-05-31 | 深圳市深网视界科技有限公司 | A kind of pedestrian based on deep learning recognition methods and device again |
| CN106778527A (en) * | 2016-11-28 | 2017-05-31 | 中通服公众信息产业股份有限公司 | A kind of improved neutral net pedestrian recognition methods again based on triple losses |
| CN107273836A (en) * | 2017-06-07 | 2017-10-20 | 深圳市深网视界科技有限公司 | A kind of pedestrian detection recognition methods, device, model and medium |
| CN107301668A (en) * | 2017-06-14 | 2017-10-27 | 成都四方伟业软件股份有限公司 | A kind of picture compression method based on sparse matrix, convolutional neural networks |
| CN107871136A (en) * | 2017-03-22 | 2018-04-03 | 中山大学 | The image-recognizing method of convolutional neural networks based on openness random pool |
| CN108197584A (en) * | 2018-01-12 | 2018-06-22 | 武汉大学 | A kind of recognition methods again of the pedestrian based on triple deep neural network |
| CN109063649A (en) * | 2018-08-03 | 2018-12-21 | 中国矿业大学 | Pedestrian's recognition methods again of residual error network is aligned based on twin pedestrian |
| CN109101913A (en) * | 2018-08-01 | 2018-12-28 | 北京飞搜科技有限公司 | Pedestrian recognition methods and device again |
| CN109344787A (en) * | 2018-10-15 | 2019-02-15 | 浙江工业大学 | A kind of specific objective tracking identified again based on recognition of face and pedestrian |
| CN109543559A (en) * | 2018-10-31 | 2019-03-29 | 东南大学 | Method for tracking target and system based on twin network and movement selection mechanism |
-
2019
- 2019-05-05 CN CN201910369005.4A patent/CN110084215A/en active Pending
Patent Citations (14)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US5778152A (en) * | 1992-10-01 | 1998-07-07 | Sony Corporation | Training method for neural network |
| CN105354568A (en) * | 2015-08-24 | 2016-02-24 | 西安电子科技大学 | Convolutional neural network based vehicle logo identification method |
| CN105608450A (en) * | 2016-03-01 | 2016-05-25 | 天津中科智能识别产业技术研究院有限公司 | Heterogeneous face identification method based on deep convolutional neural network |
| CN105975931A (en) * | 2016-05-04 | 2016-09-28 | 浙江大学 | Convolutional neural network face recognition method based on multi-scale pooling |
| CN106778464A (en) * | 2016-11-09 | 2017-05-31 | 深圳市深网视界科技有限公司 | A kind of pedestrian based on deep learning recognition methods and device again |
| CN106778527A (en) * | 2016-11-28 | 2017-05-31 | 中通服公众信息产业股份有限公司 | A kind of improved neutral net pedestrian recognition methods again based on triple losses |
| CN107871136A (en) * | 2017-03-22 | 2018-04-03 | 中山大学 | The image-recognizing method of convolutional neural networks based on openness random pool |
| CN107273836A (en) * | 2017-06-07 | 2017-10-20 | 深圳市深网视界科技有限公司 | A kind of pedestrian detection recognition methods, device, model and medium |
| CN107301668A (en) * | 2017-06-14 | 2017-10-27 | 成都四方伟业软件股份有限公司 | A kind of picture compression method based on sparse matrix, convolutional neural networks |
| CN108197584A (en) * | 2018-01-12 | 2018-06-22 | 武汉大学 | A kind of recognition methods again of the pedestrian based on triple deep neural network |
| CN109101913A (en) * | 2018-08-01 | 2018-12-28 | 北京飞搜科技有限公司 | Pedestrian recognition methods and device again |
| CN109063649A (en) * | 2018-08-03 | 2018-12-21 | 中国矿业大学 | Pedestrian's recognition methods again of residual error network is aligned based on twin pedestrian |
| CN109344787A (en) * | 2018-10-15 | 2019-02-15 | 浙江工业大学 | A kind of specific objective tracking identified again based on recognition of face and pedestrian |
| CN109543559A (en) * | 2018-10-31 | 2019-03-29 | 东南大学 | Method for tracking target and system based on twin network and movement selection mechanism |
Non-Patent Citations (2)
| Title |
|---|
| CHENG D等: "Person re-identification by multi-channel parts-based CNN with improved triplet loss function", 《PROCEEDINGS OF THE IEEE CONFERENCE ON COMPUTER VISION AND PATTERN RECOGNITION》 * |
| 罗浩等: "基于深度学习的行人重识别研究进展", 《自动化学报》 * |
Cited By (36)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110532890A (en) * | 2019-08-05 | 2019-12-03 | 安徽大学 | A kind of twin convolutional neural networks pedestrian recognition methods again of distribution based on cloud, marginal end and equipment end |
| CN110532890B (en) * | 2019-08-05 | 2021-10-22 | 安徽大学 | Distributed twin convolutional neural network pedestrian re-identification method based on cloud end, edge end and equipment end |
| CN110503053A (en) * | 2019-08-27 | 2019-11-26 | 电子科技大学 | Human motion recognition method based on cyclic convolution neural network |
| CN110503053B (en) * | 2019-08-27 | 2022-07-08 | 电子科技大学 | Human body action recognition method based on cyclic convolution neural network |
| CN110580460A (en) * | 2019-08-28 | 2019-12-17 | 西北工业大学 | Pedestrian re-identification method based on combined identification and verification of pedestrian identity and attribute characteristics |
| CN110516745A (en) * | 2019-08-28 | 2019-11-29 | 北京达佳互联信息技术有限公司 | Training method, device and the electronic equipment of image recognition model |
| CN110516745B (en) * | 2019-08-28 | 2022-05-24 | 北京达佳互联信息技术有限公司 | Training method and device of image recognition model and electronic equipment |
| CN110659585A (en) * | 2019-08-31 | 2020-01-07 | 电子科技大学 | Pedestrian detection method based on interactive attribute supervision |
| CN110659585B (en) * | 2019-08-31 | 2022-03-15 | 电子科技大学 | Pedestrian detection method based on interactive attribute supervision |
| CN110738241A (en) * | 2019-09-24 | 2020-01-31 | 中山大学 | binocular stereo vision matching method based on neural network and operation frame thereof |
| CN110837775A (en) * | 2019-09-30 | 2020-02-25 | 合肥合工安驰智能科技有限公司 | Underground locomotive pedestrian and distance detection method based on binarization network |
| CN110728330A (en) * | 2019-10-23 | 2020-01-24 | 腾讯科技（深圳）有限公司 | Object identification method, device, equipment and storage medium based on artificial intelligence |
| CN110929099A (en) * | 2019-11-28 | 2020-03-27 | 杭州趣维科技有限公司 | Short video frame semantic extraction method and system based on multitask learning |
| CN111079585A (en) * | 2019-12-03 | 2020-04-28 | 浙江工商大学 | Image enhancement and pseudo-twin convolution neural network combined pedestrian re-identification method based on deep learning |
| CN111079585B (en) * | 2019-12-03 | 2023-08-15 | 浙江工商大学 | Pedestrian re-identification method combining image enhancement with pseudo-twin convolutional neural network |
| CN111797700A (en) * | 2020-06-10 | 2020-10-20 | 南昌大学 | Vehicle re-identification method based on fine-grained discrimination network and second-order reordering |
| CN111783624B (en) * | 2020-06-29 | 2023-03-24 | 厦门市美亚柏科信息股份有限公司 | Pedestrian re-identification method and device with translation invariance reserved and storage medium |
| CN111783624A (en) * | 2020-06-29 | 2020-10-16 | 厦门市美亚柏科信息股份有限公司 | Pedestrian re-identification method and device with translation invariance reserved and storage medium |
| CN112434790A (en) * | 2020-11-10 | 2021-03-02 | 西安理工大学 | Self-interpretation method for convolutional neural network to judge partial black box problem |
| CN112434790B (en) * | 2020-11-10 | 2024-03-29 | 西安理工大学 | Self-interpretation method for distinguishing part of black box problem of convolutional neural network |
| CN112508114A (en) * | 2020-12-14 | 2021-03-16 | 武汉纺织大学 | Intelligent clothing recommendation system and method |
| CN112949384B (en) * | 2021-01-23 | 2024-03-08 | 西北工业大学 | Remote sensing image scene classification method based on antagonistic feature extraction |
| CN112949384A (en) * | 2021-01-23 | 2021-06-11 | 西北工业大学 | Remote sensing image scene classification method based on antagonistic feature extraction |
| CN113221530B (en) * | 2021-04-19 | 2024-02-13 | 杭州火石数智科技有限公司 | Text similarity matching method and device, computer equipment and storage medium |
| CN113221530A (en) * | 2021-04-19 | 2021-08-06 | 杭州火石数智科技有限公司 | Text similarity matching method and device based on circle loss, computer equipment and storage medium |
| CN113362070A (en) * | 2021-06-03 | 2021-09-07 | 中国工商银行股份有限公司 | Method, apparatus, electronic device, and medium for identifying operating user |
| CN113256592B (en) * | 2021-06-07 | 2021-10-08 | 中国人民解放军总医院 | Training method, system and device of image feature extraction model |
| CN113256592A (en) * | 2021-06-07 | 2021-08-13 | 中国人民解放军总医院 | Training method, system and device of image feature extraction model |
| CN113361693A (en) * | 2021-06-30 | 2021-09-07 | 北京百度网讯科技有限公司 | Method and device for generating convolutional neural network and image recognition method and device |
| CN113469253B (en) * | 2021-07-02 | 2024-05-14 | 河海大学 | Electric larceny detection method based on triple twinning network |
| CN113469253A (en) * | 2021-07-02 | 2021-10-01 | 河海大学 | Electricity stealing detection method based on triple twin network |
| CN113435383A (en) * | 2021-07-07 | 2021-09-24 | 中国人民解放军国防科技大学 | Remote sensing airplane target classification method and device based on double triple pseudo-twin framework |
| CN113781518A (en) * | 2021-09-10 | 2021-12-10 | 商汤集团有限公司 | Neural network structure searching method and device, electronic device and storage medium |
| CN113781518B (en) * | 2021-09-10 | 2024-05-17 | 商汤集团有限公司 | Neural network structure searching method and device, electronic equipment and storage medium |
| CN113963150B (en) * | 2021-11-16 | 2022-04-08 | 北京中电兴发科技有限公司 | Pedestrian re-identification method based on multi-scale twin cascade network |
| CN113963150A (en) * | 2021-11-16 | 2022-01-21 | 北京中电兴发科技有限公司 | Pedestrian re-identification method based on multi-scale twin cascade network |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN110084215A (en) | A kind of pedestrian of the twin network model of binaryzation triple recognition methods and system again | |
| CN108197587B (en) | Method for performing multi-mode face recognition through face depth prediction | |
| CN108182441B (en) | Parallel multichannel convolutional neural network, construction method and image feature extraction method | |
| CN109583322B (en) | Face recognition deep network training method and system | |
| CN106599797B (en) | A kind of infrared face recognition method based on local parallel neural network | |
| CN106295601B (en) | A kind of improved Safe belt detection method | |
| CN108596211B (en) | Shielded pedestrian re-identification method based on centralized learning and deep network learning | |
| CN107016413B (en) | A kind of online stage division of tobacco leaf based on deep learning algorithm | |
| Yao et al. | Robust CNN-based gait verification and identification using skeleton gait energy image | |
| CN107818302A (en) | Non-rigid multiple dimensioned object detecting method based on convolutional neural networks | |
| CN101739712A (en) | Video-based 3D human face expression cartoon driving method | |
| CN111709311A (en) | Pedestrian re-identification method based on multi-scale convolution feature fusion | |
| CN106415594A (en) | A method and a system for face verification | |
| CN104537647A (en) | Target detection method and device | |
| CN111666843A (en) | Pedestrian re-identification method based on global feature and local feature splicing | |
| CN110781928B (en) | Image similarity learning method for extracting multi-resolution features of image | |
| CN109903339B (en) | Video group figure positioning detection method based on multi-dimensional fusion features | |
| CN107977683A (en) | Joint SAR target identification methods based on convolution feature extraction and machine learning | |
| CN114419671A (en) | Hypergraph neural network-based occluded pedestrian re-identification method | |
| CN113326735B (en) | YOLOv 5-based multi-mode small target detection method | |
| CN110309770A (en) | A kind of vehicle discrimination method again based on the study of four-tuple loss metric | |
| CN110390294A (en) | Target tracking method based on bidirectional long-short term memory neural network | |
| CN104463194A (en) | Driver-vehicle classification method and device | |
| Fedorchenko et al. | Genetic Method of Image Processing for Motor Vehicle Recognition. | |
| CN109919246A (en) | Pedestrian's recognition methods again based on self-adaptive features cluster and multiple risks fusion |
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 | ||
| RJ01 | Rejection of invention patent application after publication |
Application publication date: 20190802 |