[go: up one dir, main page]

CN113095229A - Unsupervised domain self-adaptive pedestrian re-identification system and method - Google Patents

Unsupervised domain self-adaptive pedestrian re-identification system and method Download PDF

Info

Publication number
CN113095229A
CN113095229A CN202110399589.7A CN202110399589A CN113095229A CN 113095229 A CN113095229 A CN 113095229A CN 202110399589 A CN202110399589 A CN 202110399589A CN 113095229 A CN113095229 A CN 113095229A
Authority
CN
China
Prior art keywords
pedestrian
domain sample
loss function
pictures
target domain
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
Application number
CN202110399589.7A
Other languages
Chinese (zh)
Other versions
CN113095229B (en
Inventor
李佳函
李云龙
程德强
寇旗旗
张皓翔
韩成功
徐进洋
江曼
刘瑞航
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
China University of Mining and Technology CUMT
Original Assignee
Jiangsu Huatu Mining Technology Co ltd
China University of Mining and Technology CUMT
Priority date (The priority date 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 date listed.)
Filing date
Publication date
Application filed by Jiangsu Huatu Mining Technology Co ltd, China University of Mining and Technology CUMT filed Critical Jiangsu Huatu Mining Technology Co ltd
Priority to CN202110399589.7A priority Critical patent/CN113095229B/en
Publication of CN113095229A publication Critical patent/CN113095229A/en
Application granted granted Critical
Publication of CN113095229B publication Critical patent/CN113095229B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Classifications

    • 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
    • G06V40/103—Static body considered as a whole, e.g. static pedestrian or occupant recognition
    • 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
    • G06F18/22—Matching criteria, e.g. proximity measures
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/24—Classification techniques
    • G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V20/00—Scenes; Scene-specific elements
    • G06V20/40—Scenes; Scene-specific elements in video content
    • G06V20/46—Extracting features or characteristics from the video content, e.g. video fingerprints, representative shots or key frames

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Data Mining & Analysis (AREA)
  • Physics & Mathematics (AREA)
  • General Physics & Mathematics (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Evolutionary Biology (AREA)
  • Evolutionary Computation (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • General Engineering & Computer Science (AREA)
  • Artificial Intelligence (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Multimedia (AREA)
  • Human Computer Interaction (AREA)
  • Image Analysis (AREA)

Abstract

The invention relates to an unsupervised domain self-adaptive pedestrian re-identification system and method, belongs to the technical field of pedestrian re-identification, and solves the problems of high difficulty and low identification accuracy rate of the existing unsupervised domain self-adaptive pedestrian re-identification. The system comprises a data acquisition module, a data acquisition module and a data processing module, wherein the data acquisition module is used for acquiring a plurality of source domain sample subsets and a plurality of target domain sample subsets; the network model training module is used for obtaining a classification loss function and a sample invariance loss function of the pedestrian re-recognition network model, sequencing and layering pedestrian pictures in the target domain sample subset according to the similarity between each pedestrian picture in the source domain sample subset and each pedestrian picture in the target domain sample subset to obtain a layering loss function, and further performing iterative optimization on the pedestrian re-recognition network model; and the re-identification module is used for re-identifying the pedestrians by utilizing the optimized pedestrian re-identification network model to obtain the images which are the same as or similar to the images of the pedestrians to be identified. The system can reduce the migration loss of the network and improve the accuracy of pedestrian re-identification.

Description

Unsupervised domain self-adaptive pedestrian re-identification system and method
Technical Field
The invention relates to the technical field of pedestrian re-identification, in particular to an unsupervised domain self-adaptive pedestrian re-identification system and method.
Background
In conventional unsupervised domain adaptive learning, most methods are used in a closed set scenario where sample classes are consistent in the source domain as well as in the target domain. However, the conventional unsupervised domain adaptive algorithm cannot be used for unsupervised domain adaptive pedestrian re-recognition, because due to image acquisition, the pedestrian categories of the source domain and the target domain are almost different, and a large migration loss is generated when a network trained by using source domain images migrates to the target domain for recognition. The large migration loss caused by network migration makes unsupervised domain adaptive pedestrian re-identification more challenging than most unsupervised learning.
In the prior art, unsupervised domain self-adaptive pedestrian re-identification mainly achieves the purpose of distinguishing label-free data sets through self-clustering of target domains. In the prior art, three invariance losses based on a target domain are proposed to reduce the migration loss of the network, but marked data are not fully utilized, and deep semantic information of a non-tag picture is not fully mined, so that the migration loss of the network cannot be efficiently reduced, and the accuracy of pedestrian re-identification needs to be improved.
In the prior art, at least the following defects exist, the source domain picture information and the target domain picture information cannot be fully utilized, and the network distinguishing learning cannot be carried out on negative samples with different similarities according to the similarities between the source domain picture and the target domain picture, so that the migration loss of a pedestrian re-identification network is large, and the accuracy of the result of the pedestrian re-identification is low.
Disclosure of Invention
In view of the foregoing analysis, embodiments of the present invention provide an unsupervised domain adaptive pedestrian re-identification system and method, so as to solve the problems of large migration loss of a pedestrian re-identification network and low accuracy of a result of pedestrian re-identification in the prior art.
In one aspect, the present invention provides an unsupervised domain adaptive pedestrian re-identification system, including:
a data acquisition module for acquiring a source domain sample set comprising a plurality of source domain sample subsets and a target domain sample set comprising a plurality of target domain sample subsets;
the network model training module is used for obtaining a classification loss function and a sample invariance loss function of the pedestrian re-recognition network model, sequencing and layering pedestrian pictures in the target domain sample subset according to the similarity between each pedestrian picture in the source domain sample subset and each pedestrian picture in the target domain sample subset, and obtaining a layering loss function based on layered pedestrian picture characteristics and corresponding layering weights; traversing each source domain sample subset and each target domain sample subset based on the classification loss function, the sample invariance loss function and the hierarchical loss function to perform iterative optimization on the pedestrian re-identification network model;
and the re-identification module is used for identifying the pedestrian image to be identified by utilizing the optimized pedestrian re-identification network model to obtain the image which is the same as or similar to the pedestrian image to be identified.
Further, the pedestrian re-identification network model comprises a residual error network structure, a full connection layer and a Softmax normalization layer which are connected in sequence and correspond to the classification loss function, a target domain memory and an L, wherein the target domain memory and the L are connected in sequence and correspond to the sample invariance loss function2The system comprises a normalization layer, a source domain memory and a similarity measurement axis network structure which are sequentially connected and correspond to a hierarchical loss function, wherein the residual error network structure is respectively connected with the full connection layer, the target domain memory, the source domain memory and the similarity measurement axis network structure.
Further, the network model training module obtains the hierarchical loss function specifically by:
respectively inputting the source domain sample subset and the target domain sample subset into a residual error network structure of a pedestrian re-identification network model to extract image features so as to obtain and respectively store the features of each pedestrian picture in the source domain sample subset and the features of each pedestrian picture in the target domain sample subset, and multiplying the features of each pedestrian picture in the source domain sample subset and the features of each pedestrian picture in the target domain sample subset to obtain corresponding similarity;
based on the similarity, sequencing the pedestrian pictures in the target domain sample subset in a descending order, sequentially selecting a first preset number of pictures as first-layer pictures, selecting a second preset number of pictures as second-layer pictures, taking the rest pictures as third-layer pictures, and respectively setting layer weights of the three layers of pictures;
obtaining a hierarchical loss function based on the characteristics of each pedestrian picture in the target domain sample subset and the level weight corresponding to the level to which the pedestrian picture belongs:
Figure BDA0003019905840000031
Figure BDA0003019905840000032
wherein L isSLRepresenting the hierarchical loss function, ntRepresenting the number of pedestrian pictures, w, in the target field sample subsett,rRepresenting the level weight, r representing the sequence number of the pedestrian pictures in the target domain sample subset according to the similarity, xt,iRepresenting target domain sample subset input pedestrian re-recognitionInputting a pedestrian picture with the sequence of i, f (x) in the network modelt,i) Picture x representing pedestriant,iP (m | x)t,i) Picture x representing pedestriant,iProbability of similarity to all classes of pedestrian pictures in all stored subsets of source domain samples, R1[m]Features, R, representing all classes of pedestrian pictures in all stored subsets of active domain samples1[j]Features representing class j pedestrian pictures in all stored subsets of active domain samples, NsAnd the stored category number of the pedestrian pictures in all the active domain sample subsets is represented, and beta is a temperature coefficient.
Further, the layer weights of the three layers of pictures are set as follows:
Figure BDA0003019905840000041
Figure BDA0003019905840000042
where b denotes the number of first layer pictures, k1Represents the sum of the numbers of the first layer picture and the second layer picture, and p (l | i) represents the pedestrian picture xt,iProbability of class i pedestrian pictures in all target domain sample subsets, R1[l]Features representing class i pedestrian pictures in all stored subsets of target domain samples, NtRepresenting the number of categories of pedestrian pictures in all the stored subsets of the target domain samples.
Further, the network model training module obtains a classification loss function specifically by the following means:
inputting the source domain sample subset into a residual error network structure of a pedestrian re-identification network model to extract image features so as to obtain the features of each pedestrian picture in the source domain sample subset;
sequentially inputting the characteristics of each pedestrian picture into a full connection layer and a softmax regression layer of a pedestrian re-identification network model, and performing characteristic dimension conversion and characteristic normalization;
classifying a loss function by adopting the following formula based on the characteristics of each pedestrian picture after dimension conversion and normalization;
Figure BDA0003019905840000043
Figure BDA0003019905840000044
wherein L issrcRepresenting the classification loss function, nsRepresenting the number of pedestrian pictures, x, in a subset of source domain sampless,mRepresents the m-th pedestrian picture in the source domain sample subset, f (x)s,m) Picture x representing pedestrians,mCharacteristic of (a), ys,mRepresenting a pedestrian picture x in a subset of source domain sampless,mClass label of p (y)s,m|xs,m) Picture x representing pedestrians,mBelong to the category ys,mProbability of (a), f (x)s,j) Representing features of class j pedestrian pictures in the source domain sample subset, NsRepresenting the number of categories of pedestrian pictures in all the stored subsets of the active domain samples.
Further, the network model training module obtains a sample invariance loss function specifically by the following means:
inputting the target domain sample subset into a residual error network structure of a pedestrian re-identification network model to extract image features so as to obtain the features of each pedestrian picture in the target domain sample subset, and performing L on the features of each pedestrian picture2Carrying out standardization treatment;
obtaining a sample invariance loss function according to the characteristics of each pedestrian picture after the normalization processing through the following formula:
Figure BDA0003019905840000051
Figure BDA0003019905840000052
wherein L isTRepresenting the sample invariance loss function, ntRepresenting the number, x, of pictures of a person in a sample subset of the target fieldt,iRepresenting the pedestrian picture with the input sequence i when the target domain sample subset is input into the pedestrian re-identification network model, f (x)t,i) Picture x representing pedestriant,iIs characterized by wi,lPicture x representing pedestriant,iWeight of pedestrian pictures belonging to class I, p (l | x)t,i) Picture x representing pedestriant,iProbability of a pedestrian belonging to class I picture, NtAnd the stored category number of the pedestrian pictures in all the target domain sample subsets is represented, and beta is a temperature coefficient.
Further, the network model training module performs iterative optimization on the pedestrian re-recognition network model specifically by the following means:
obtaining a total loss function according to the classification loss function, the sample invariance loss function and the layering loss function:
L＝λ1Lsrc+λ2LT+λ3LSL，
wherein L represents the total loss function, LsrcRepresenting the classification loss function, λ1Weight representing the classification loss function, LTRepresenting the sample invariance loss function, λ2Weight, L, representing the sample invariance loss functionSLRepresenting the hierarchical loss function, λ3A weight representing a hierarchical loss function;
and traversing each source domain sample subset and each target domain sample subset, and iteratively updating the total loss function until the variable quantity of the total loss function value is smaller than a preset value so as to complete the optimization of the pedestrian re-identification network model.
On the other hand, the invention provides an unsupervised domain self-adaptive pedestrian re-identification method, which comprises the following steps:
obtaining a source domain sample set comprising a plurality of source domain sample subsets and a target domain sample set comprising a plurality of target domain sample subsets;
respectively obtaining a classification loss function and a sample invariance loss function of a pedestrian re-identification network model based on the source domain sample subset and the target domain sample subset, sequencing and layering pedestrian pictures in the target domain sample subset according to the similarity of each pedestrian picture in the source domain sample subset and each pedestrian picture in the target domain sample subset, and obtaining a layering loss function based on layered pedestrian picture characteristics and corresponding layering weights; traversing each source domain sample subset and each target domain sample subset based on the classification loss function, the sample invariance loss function and the hierarchical loss function to perform iterative optimization on the pedestrian re-identification network model;
and identifying the image of the pedestrian to be identified by utilizing the optimized pedestrian re-identification network model to obtain the image which is the same as or similar to the image of the pedestrian to be identified.
Further, the hierarchical loss function is obtained by specifically:
respectively inputting the source domain sample subset and the target domain sample subset into a residual error network structure of a pedestrian re-identification network model to extract image features so as to obtain and respectively store the features of each pedestrian picture in the source domain sample subset and the features of each pedestrian picture in the target domain sample subset, and multiplying the features of each pedestrian picture in the source domain sample subset and the features of each pedestrian picture in the target domain sample subset to obtain corresponding similarity;
based on the similarity, sequencing the pedestrian pictures in the target domain sample subset in a descending order, sequentially selecting a first preset number of pictures as first-layer pictures, selecting a second preset number of pictures as second-layer pictures, taking the rest pictures as third-layer pictures, and respectively setting layer weights of the three layers of pictures;
obtaining a hierarchical loss function based on the characteristics of each pedestrian picture in the target domain sample subset and the level weight corresponding to the level to which the pedestrian picture belongs:
Figure BDA0003019905840000071
Figure BDA0003019905840000072
wherein L isSLRepresenting the hierarchical loss function, ntRepresenting the number of pedestrian pictures, w, in the target field sample subsett,rRepresenting the level weight, r representing the sequence number of the pedestrian pictures in the target domain sample subset according to the similarity, xt,iRepresenting the pedestrian picture with the input sequence i when the target domain sample subset is input into the pedestrian re-identification network model, f (x)t,i) Picture x representing pedestriant,iP (m | x)t,i) Picture x representing pedestriant,iProbability of similarity to all classes of pedestrian pictures in all stored subsets of source domain samples, R1[m]Features, R, representing all classes of pedestrian pictures in all stored subsets of active domain samples1[j]Features representing class j pedestrian pictures in all stored subsets of active domain samples, NsAnd the stored category number of the pedestrian pictures in all the active domain sample subsets is represented, and beta is a temperature coefficient.
Further, the layer weights of the three layers of pictures are set as follows:
Figure BDA0003019905840000073
Figure BDA0003019905840000074
where b denotes the number of first layer pictures, k1Represents the sum of the numbers of the first layer picture and the second layer picture, and p (l | i) represents the pedestrian picture xt,iProbability of class i pedestrian pictures in all target domain sample subsets, R1[l]Features representing class i pedestrian pictures in all stored subsets of target domain samples, NtRepresenting the number of categories of pedestrian pictures in all the stored subsets of the target domain samples.
Compared with the prior art, the invention can realize at least one of the following beneficial effects:
1. the unsupervised domain self-adaptive pedestrian re-identification system and the unsupervised domain self-adaptive pedestrian re-identification method effectively combine the supervised pedestrian re-identification with the unsupervised pedestrian re-identification through the combination of the source domain sample with the label and the target domain sample without the label, and firstly propose to layer the target domain sample through the similarity between the source domain sample and the target domain sample, so that the discrimination of the pedestrian re-identification network on the target domain sample is improved, the migration loss of the network is reduced, and the accuracy of the pedestrian re-identification result of the pedestrian re-identification network is further improved.
2. The unsupervised domain self-adaptive pedestrian re-recognition system and the unsupervised domain self-adaptive pedestrian re-recognition method, provided by the invention, utilize the similarity of the characteristics of the source domain sample and the target domain sample to carry out layering and weighting on the negative sample to obtain a layered loss function, the loss function enables the network to be self-adaptive in the training of a pedestrian re-recognition network model so as to reduce the learning of the characteristics of the negative sample, thereby reducing the migration loss of the network, and on the basis, the classification loss function and the sample invariance loss function are combined to carry out iterative optimization on the pedestrian re-recognition network, so that the accuracy of the pedestrian re-recognition result of the pedestrian re-recognition network.
In the invention, the technical schemes can be combined with each other to realize more preferable combination schemes. Additional features and advantages of the invention will be set forth in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. The objectives and other advantages of the invention will be realized and attained by the structure particularly pointed out in the written description and drawings.
Drawings
The drawings are only for purposes of illustrating particular embodiments and are not to be construed as limiting the invention, wherein like reference numerals are used to designate like parts throughout.
FIG. 1 is a schematic diagram of an unsupervised domain adaptive pedestrian re-identification system according to an embodiment of the present invention;
FIG. 2 is a diagram illustrating a pedestrian re-identification network model according to an embodiment of the present invention;
fig. 3 is a flowchart of an unsupervised domain adaptive pedestrian re-identification method according to an embodiment of the present invention.
Reference numerals:
110-a data acquisition module; 120-a network model training module; 130-re-identification module.
Detailed Description
The accompanying drawings, which are incorporated in and constitute a part of this application, illustrate preferred embodiments of the invention and together with the description, serve to explain the principles of the invention and not to limit the scope of the invention.
System embodiment
The invention discloses an unsupervised domain self-adaptive pedestrian re-identification system. As shown in fig. 1, the system includes:
a data acquisition module 110 for acquiring a source domain sample set comprising a plurality of source domain sample subsets and a target domain sample set comprising a plurality of target domain sample subsets. Specifically, a large number of pictures of pedestrians are randomly acquired from different angles, each picture contains a pedestrian, the acquired pictures are divided into a training set and a testing set, the training set is used for training the pedestrian re-recognition network model, and the testing set is used for testing the trained pedestrian re-recognition network model so as to ensure the recognition accuracy of the pedestrian re-recognition network model. The method comprises the steps of randomly selecting partial pictures in a training set, adding a label to each picture, specifically, adding the same label to a plurality of pictures of the same pedestrian, wherein different labels represent different pedestrians, dividing the pictures with the labels added into a plurality of source domain sample subsets, taking other pictures without the labels in the training set as a target domain sample set, adding a number to each picture, and randomly dividing the pictures with the numbers added into a plurality of target domain sample subsets.
The network model training module 120 is configured to obtain a classification loss function and a sample invariance loss function of the pedestrian re-recognition network model, rank and layer pedestrian pictures in the target domain sample subset according to a similarity between each pedestrian picture in the source domain sample subset and each pedestrian picture in the target domain sample subset, and obtain a layer loss function based on layered pedestrian picture features and corresponding layer weights. And traversing each source domain sample subset and each target domain sample subset based on a classification loss function, a sample invariance loss function and a hierarchical loss function to perform iterative optimization on the pedestrian re-identification network model.
And the re-identification module is used for identifying the pedestrian image to be identified by utilizing the optimized pedestrian re-identification network model to obtain the image which is the same as or similar to the pedestrian image to be identified. Specifically, a to-be-identified pedestrian picture is input into a trained pedestrian re-identification network model, the model can output a picture most similar to the to-be-identified pedestrian picture, exemplarily, the model is set to output the number of the first three pictures with the highest similarity to the to-be-identified pedestrian picture, the first three pictures are pictures in a target domain sample set, and then the to-be-identified pedestrian picture and the first three pictures are manually compared to determine the category of the to-be-identified pedestrian picture.
Preferably, as shown in fig. 2, the pedestrian re-identification network model includes a residual network structure, a fully-connected layer and a Softmax normalization layer connected in sequence corresponding to the classification loss function, a target domain memory connected in sequence corresponding to the sample invariance loss function, and an L2The system comprises a normalization layer, a source domain memory and a similarity measurement axis network structure which are connected in sequence and correspond to a hierarchical loss function, wherein the residual error network structure is connected with a full connection layer, a target domain memory, the source domain memory and the similarity measurement axis network structure respectively. Specifically, the target domain memory and the source domain memory are both in a key value storage structure, the key is used for storing the characteristics of the pedestrian picture, the value is used for storing a number or a label corresponding to the pedestrian picture, and the residual network structure is ResNet 50.
Preferably, the network model training module obtains the classification loss function, the sample invariance loss function and the hierarchical loss function in the following way:
step 1, inputting the source domain sample subset and the target domain sample subset into a residual error network structure of a pedestrian re-identification network model respectively to extract image features so as to obtain the features of each pedestrian picture in the source domain sample subset and the features of each pedestrian picture in the target domain sample subset.
And storing the characteristics of each pedestrian picture in the target domain sample subset in a target domain memory, and storing the characteristics of each pedestrian picture in the source domain sample subset in a source domain memory.
Step 2, specifically, obtaining a classification loss function in the following way:
and sequentially inputting the characteristics of each pedestrian picture in the source domain sample subset into a full connection layer and a softmax regression layer of the pedestrian re-identification network model, and performing characteristic dimension conversion and characteristic normalization to strengthen the nonlinearity of the pedestrian re-identification network model in learning.
Classifying a loss function by adopting the following formula based on the characteristics of each pedestrian picture after dimension conversion and normalization;
Figure BDA0003019905840000111
Figure BDA0003019905840000112
wherein L issrcRepresenting the classification loss function, nsRepresenting the number of pedestrian pictures, x, in a subset of source domain sampless,mRepresents the m-th pedestrian picture in the source domain sample subset, f (x)s,m) Picture x representing pedestrians,mCharacteristic of (a), ys,mRepresenting a pedestrian picture x in a subset of source domain sampless,mClass label of p (y)s,m|xs,m) Picture x representing pedestrians,mBelong to the category ys,mProbability of (a), f (x)s,j) Representing features of class j pedestrian pictures in the source domain sample subset, NsRepresenting the number of categories of pedestrian pictures in all the stored subsets of the active domain samples.
Step 3, specifically, obtaining a sample invariance loss function through the following method:
l is carried out on the characteristics of each pedestrian picture in the current target domain sample subset stored in the target domain memory2And (6) carrying out standardization processing.
Obtaining a sample invariance loss function according to the characteristics of each pedestrian picture after the normalization processing through the following formula:
Figure BDA0003019905840000121
Figure BDA0003019905840000122
wherein L isTRepresenting the sample invariance loss function, ntRepresenting the number, x, of pictures of a person in the current target field sample subsett,iRepresenting the pedestrian picture with the input sequence i when the current target domain sample subset is input into the pedestrian re-identification network model, f (x)t,i) Picture x representing pedestriant,iIs characterized by wi,lPicture x representing pedestriant,iWeight of pedestrian pictures belonging to class I, p (l | x)t,i) Picture x representing pedestriant,iProbability of a pedestrian belonging to class I picture, NtAnd the stored category number of the pedestrian pictures in all the target domain sample subsets is represented, and beta is a temperature coefficient.
Step 4, obtaining a layering loss function by the following method:
multiplying the characteristics of each pedestrian picture in the current source domain sample subset and the characteristics of each pedestrian picture in the current target domain sample subset based on the similarity measurement axis network structure to obtain corresponding similarity;
and sequencing the pedestrian pictures in the target domain sample subset in a descending order based on the similarity, sequentially selecting a first preset number of pictures as first-layer pictures, selecting a second preset number of pictures as second-layer pictures, taking the rest pictures as third-layer pictures, and respectively setting layer weights of the three layers of pictures. Illustratively, the number of first layer pictures is 3, and the number of second layer pictures is 147.
Obtaining a hierarchical loss function based on the characteristics of each pedestrian picture in the target domain sample subset and the level weight corresponding to the level to which the pedestrian picture belongs:
Figure BDA0003019905840000123
Figure BDA0003019905840000131
wherein L isSLRepresenting the hierarchical loss function, ntRepresenting the number of pedestrian pictures, w, in the target field sample subsett,rRepresenting the level weight, r representing the sequence number of the pedestrian pictures in the target domain sample subset according to the similarity, xt,iRepresenting the pedestrian picture with the input sequence i when the target domain sample subset is input into the pedestrian re-identification network model, f (x)t,i) Picture x representing pedestriant,iP (m | x)t,i) Picture x representing pedestriant,iProbability of similarity to all classes of pedestrian pictures in all stored subsets of source domain samples, R1[m]Features, R, representing all classes of pedestrian pictures in all stored subsets of active domain samples1[j]Features representing class j pedestrian pictures in all stored subsets of active domain samples, NsAnd the stored category number of the pedestrian pictures in all the active domain sample subsets is represented, and beta is a temperature coefficient.
Preferably, the layer weights of the three layers of pictures are set as follows:
Figure BDA0003019905840000132
Figure BDA0003019905840000133
where b denotes the number of first layer pictures, k1Represents the sum of the numbers of the first layer picture and the second layer picture, and p (l | i) represents the pedestrian picture xt,iProbability of class i pedestrian pictures in all target domain sample subsets, R1[l]Features representing class i pedestrian pictures in all stored subsets of target domain samples, NtRepresenting pedestrians in all stored subsets of target domain samplesNumber of categories of pictures.
Specifically, the steps 2 to 4 are not sequentially performed, and can be performed simultaneously.
Step 5, specifically, the pedestrian re-identification network model is subjected to iterative optimization in the following mode:
obtaining a total loss function according to the classification loss function, the sample invariance loss function and the layering loss function:
L＝λ1Lsrc+λ2LT+λ3LSL，
wherein L represents the total loss function, LsrcRepresenting the classification loss function, λ1Weight representing the classification loss function, LTRepresenting the sample invariance loss function, λ2Weight, L, representing the sample invariance loss functionSLRepresenting the hierarchical loss function, λ3Weights representing a hierarchical loss function, exemplary, λ1A value of 0.7, λ2A value of 0.3, λ3Is 0.2.
And traversing each source domain sample subset and each target domain sample subset, and repeating the steps 1 to 5 to update the total loss function in an iterative manner until the variable quantity of the total loss function value is smaller than a preset value so as to complete the optimization of the pedestrian re-identification network model.
Specifically, in the iterative update process, the pedestrian image features of the target domain sample subset stored in the target domain memory are updated in real time in the following manner:
Figure BDA0003019905840000141
xt,irepresenting the pedestrian picture with the input sequence i when the target domain sample subset is input into the pedestrian re-identification network model, f (x)t,i) Picture x representing pedestriant,iWarp L2Normalized characteristic, RiImage x representing pedestrian in target domain stored in target domain memoryt,iIs characterized in that it is a mixture of two or more of the above-mentioned components,
Figure BDA0003019905840000142
is a hyper-parameter that controls the speed of feature updates.
Method embodiment
The embodiment of the method is based on the same principle as the system embodiment, so that the method is not repeated herein, and repeated parts can refer to the system embodiment on the market.
Specifically, as shown in fig. 3, the method includes:
s110, obtaining a source domain sample set comprising a plurality of source domain sample subsets and a target domain sample set comprising a plurality of target domain sample subsets.
S120, respectively obtaining a classification loss function and a sample invariance loss function of the pedestrian re-identification network model based on the source domain sample subset and the target domain sample subset, sequencing and layering pedestrian pictures in the target domain sample subset according to the similarity of each pedestrian picture in the source domain sample subset and each pedestrian picture in the target domain sample subset, and obtaining a layering loss function based on layered pedestrian picture characteristics and corresponding layering weights; and traversing each source domain sample subset and each target domain sample subset based on the classification loss function, the sample invariance loss function and the hierarchical loss function to perform iterative optimization on the pedestrian re-identification network model.
S130, identifying the pedestrian image to be identified by utilizing the optimized pedestrian re-identification network model, and obtaining the image which is the same as or similar to the pedestrian image to be identified.
Preferably, the stratification loss function is obtained in particular by:
respectively inputting the source domain sample subset and the target domain sample subset into a residual error network structure of a pedestrian re-identification network model to extract image features so as to obtain and respectively store the features of each pedestrian picture in the source domain sample subset and the features of each pedestrian picture in the target domain sample subset, and multiplying the features of each pedestrian picture in the source domain sample subset and the features of each pedestrian picture in the target domain sample subset to obtain corresponding similarity;
based on the similarity, sequencing the pedestrian pictures in the target domain sample subset in a descending order, sequentially selecting a first preset number of pictures as first-layer pictures, selecting a second preset number of pictures as second-layer pictures, taking the rest pictures as third-layer pictures, and respectively setting layer weights of the three layers of pictures;
obtaining a hierarchical loss function based on the characteristics of each pedestrian picture in the target domain sample subset and the level weight corresponding to the level to which the pedestrian picture belongs:
Figure BDA0003019905840000151
Figure BDA0003019905840000152
wherein L isSLRepresenting the hierarchical loss function, ntRepresenting the number of pedestrian pictures, w, in the target field sample subsett,rRepresenting the level weight, r representing the sequence number of the pedestrian pictures in the target domain sample subset according to the similarity, xt,iRepresenting the pedestrian picture with the input sequence i when the target domain sample subset is input into the pedestrian re-identification network model, f (x)t,i) Picture x representing pedestriant,iP (m | x)t,i) Picture x representing pedestriant,iProbability of similarity to all classes of pedestrian pictures in all stored subsets of source domain samples, R1[m]Features, R, representing all classes of pedestrian pictures in all stored subsets of active domain samples1[j]Features representing class j pedestrian pictures in all stored subsets of active domain samples, NsAnd the stored category number of the pedestrian pictures in all the active domain sample subsets is represented, and beta is a temperature coefficient.
Preferably, the layer weights of the three layers of pictures are set as follows:
Figure BDA0003019905840000161
Figure BDA0003019905840000162
where b denotes the number of first layer pictures, k1Represents the sum of the numbers of the first layer picture and the second layer picture, and p (l | i) represents the pedestrian picture xt,iProbability of class i pedestrian pictures in all target domain sample subsets, R1[l]Features representing class i pedestrian pictures in all stored subsets of target domain samples, NtRepresenting the number of categories of pedestrian pictures in all the stored subsets of the target domain samples.
Compared with the prior art, the unsupervised domain self-adaptive pedestrian re-identification system and the unsupervised domain self-adaptive pedestrian re-identification method provided by the embodiment have the advantages that on one hand, the combination of the source domain sample with the label and the target domain sample without the label effectively combines the supervised pedestrian re-identification with the unsupervised pedestrian re-identification, namely, the target domain sample is layered through the similarity between the source domain sample and the target domain sample for the first time, so that the discrimination of the pedestrian re-identification network on the target domain sample is improved, the migration loss of the network is reduced, and the accuracy of the pedestrian re-identification result by the pedestrian re-identification network is improved; on the other hand, the negative samples are layered and weighted by utilizing the similarity of the characteristics of the source domain samples and the target domain samples to obtain a layered loss function, the loss function enables the network to adaptively reduce the learning of the characteristics of the negative samples in the training of the pedestrian re-recognition network model, so that the migration loss of the network is reduced, on the basis, the classification loss function and the sample invariance loss function are combined to perform iterative optimization on the pedestrian re-recognition network, and the accuracy of the pedestrian re-recognition network on the pedestrian re-recognition result can be greatly improved.
Those skilled in the art will appreciate that all or part of the flow of the method implementing the above embodiments may be implemented by a computer program, which is stored in a computer readable storage medium, to instruct related hardware. The computer readable storage medium is a magnetic disk, an optical disk, a read-only memory or a random access memory.
The above description is only for the preferred embodiment of the present invention, but the scope of the present invention is not limited thereto, and any changes or substitutions that can be easily conceived by those skilled in the art within the technical scope of the present invention are included in the scope of the present invention.

Claims (10)

1.一种无监督域自适应行人重识别系统，其特征在于，包括：1. an unsupervised domain adaptive pedestrian re-identification system, is characterized in that, comprises: 数据获取模块，用于获取包括多个源域样本子集的源域样本集和包括多个目标域样本子集的目标域样本集；a data acquisition module for acquiring a source domain sample set including multiple source domain sample subsets and a target domain sample set including multiple target domain sample subsets; 网络模型训练模块，用于获得行人重识别网络模型的分类损失函数和样本不变性损失函数，以及根据所述源域样本子集中每一行人图片与目标域样本子集中每一行人图片的相似度对目标域样本子集中的行人图片进行排序、分层，基于分层后的行人图片特征以及对应的层别权重获得分层损失函数；基于所述分类损失函数、样本不变性损失函数以及分层损失函数遍历每一源域样本子集和目标域样本子集对所述行人重识别网络模型进行迭代优化；The network model training module is used to obtain the classification loss function and sample invariance loss function of the pedestrian re-identification network model, and according to the similarity between each pedestrian image in the source domain sample subset and each pedestrian image in the target domain sample subset Sort and stratify the pedestrian pictures in the target domain sample subset, and obtain a hierarchical loss function based on the stratified pedestrian picture features and corresponding layer weights; based on the classification loss function, the sample invariance loss function and the hierarchical loss function The loss function traverses each source domain sample subset and target domain sample subset to iteratively optimize the pedestrian re-identification network model; 重识别模块，用于利用优化好的行人重识别网络模型对待识别行人图像进行识别，获得与所述待识别行人图像相同或相似的图像。The re-identification module is used for recognizing the image of the pedestrian to be recognized by using the optimized pedestrian re-recognition network model to obtain an image identical or similar to the image of the pedestrian to be recognized. 2.根据权利要求1所述的无监督域自适应行人重识别方法，其特征在于，所述行人重识别网络模型包括残差网络结构、与分类损失函数相对应的依次连接的全连接层和Softmax归一化层、与样本不变性损失函数对应的依次连接的目标域存储器、L2标准化层，以及与分层损失函数对应的依次连接的源域存储器和相似度衡量轴网络结构，所述残差网络结构分别与所述全连接层、目标域存储器、源域存储器以及相似度衡量轴网络结构连接。2. The unsupervised domain adaptive pedestrian re-identification method according to claim 1, wherein the pedestrian re-identification network model comprises a residual network structure, a sequentially connected fully connected layer corresponding to a classification loss function and a Softmax normalization layer, sequentially connected target domain memory corresponding to sample invariance loss function, L2 normalization layer, and sequentially connected source domain memory and similarity measure axis network structure corresponding to hierarchical loss function, said The residual network structure is respectively connected with the fully connected layer, the target domain memory, the source domain memory and the similarity measure axis network structure. 3.根据权利要求1或2所述的无监督域自适应行人重识别方法，其特征在于，所述网络模型训练模块具体通过下述方式获得分层损失函数：3. The unsupervised domain adaptive pedestrian re-identification method according to claim 1 or 2, wherein the network model training module specifically obtains a layered loss function in the following manner: 将源域样本子集和目标域样本子集分别输入行人重识别网络模型的残差网络结构中提取图像特征，以获取源域样本子集中每一行人图片的特征和目标域样本子集中每一行人图片的特征，并分别进行存储，将源域样本子集中每一行人图片的特征与目标域样本子集中每一行人图片的特征相乘获得对应的相似度；The source domain sample subset and the target domain sample subset are respectively input into the residual network structure of the person re-identification network model to extract image features to obtain the features of each pedestrian image in the source domain sample subset and each row in the target domain sample subset. The features of the person pictures are stored separately, and the features of each pedestrian picture in the source domain sample subset are multiplied by the characteristics of each pedestrian picture in the target domain sample subset to obtain the corresponding similarity; 基于所述相似度将目标域样本子集中的行人图片进行降序排序，并按顺序依次选取第一预设数量的图片作为第一层图片，选取第二预设数量的图片作为第二层图片，其余图片作为第三层图片，并分别设置三层图片的层别权重；Based on the similarity, the pedestrian pictures in the target domain sample subset are sorted in descending order, and the first preset number of pictures are selected as the first layer pictures in sequence, and the second preset number of pictures are selected as the second layer pictures, The rest of the pictures are used as the third-layer pictures, and the layer weights of the three-layer pictures are set respectively; 基于目标域样本子集中每一行人图片的特征以及其所属层别对应的层别权重获得分层损失函数：The hierarchical loss function is obtained based on the characteristics of each pedestrian image in the target domain sample subset and the layer weight corresponding to the layer to which it belongs:
Figure FDA0003019905830000021
Figure FDA0003019905830000021
Figure FDA0003019905830000022
Figure FDA0003019905830000022
 其中，LSL表示分层损失函数，nt表示目标域样本子集中行人图片的数量，wt,r表示层别权重，r表示目标域样本子集中按照相似度排序的行人图片的排序序号，xt,i表示目标域样本子集输入行人重识别网络模型时输入顺序为i的行人图片，f(xt,i)表示行人图片xt,i的特征，p(m|xt,i)表示行人图片xt,i与存储的所有源域样本子集中所有类别的行人图片的相似概率，R1[m]表示存储的所有源域样本子集中所有类别的行人图片的特征，R1[j]表示存储的所有源域样本子集中第j类行人图片的特征，Ns表示存储的所有源域样本子集中行人图片的类别数，β为温度系数。Among them, L SL represents the layered loss function, n t represents the number of pedestrian images in the target domain sample subset, w t, r represents the layer weight, r represents the sequence number of the pedestrian images sorted by similarity in the target domain sample subset, x t,i represents the target domain sample subset inputting the pedestrian image in the order of i when inputting the pedestrian re-identification network model, f(x t, i ) represents the feature of the pedestrian image x t, i , p(m|x t, i ) represents the similarity probability between the pedestrian image x t,i and the stored pedestrian images of all categories in all source domain sample subsets, R 1 [m] represents the stored features of all categories of pedestrian images in all source domain sample subsets, R 1 [j] represents the characteristics of the j-th pedestrian image in all the source domain sample subsets stored, N s represents the number of categories of pedestrian images in all the source domain sample subsets stored, and β is the temperature coefficient. 4.根据权利要求3所述的无监督域自适应行人重识别方法，其特征在于，设置的所述三层图片的层别权重为：4. The unsupervised domain adaptive pedestrian re-identification method according to claim 3, wherein the layer weights of the three-layer pictures set are:
Figure FDA0003019905830000023
Figure FDA0003019905830000023
Figure FDA0003019905830000031
Figure FDA0003019905830000031
 其中，b表示第一层图片的数量，k1表示第一层图片和第二层图片的数量之和，p(l|i)表示行人图片xt,i属于存储的所有目标域样本子集中第l类行人图片的概率，R1[l]表示存储的所有目标域样本子集中第l类行人图片的特征，Nt表示存储的所有目标域样本子集中行人图片的类别数。Among them, b represents the number of pictures in the first layer, k 1 represents the sum of the number of pictures in the first layer and the pictures in the second layer, p(l|i) represents the pedestrian picture x t, i belongs to all the target domain sample subsets stored The probability of the l-th type of pedestrian image, R 1 [l] represents the characteristics of the l-th type of pedestrian image in all the target domain sample subsets stored, and N t represents the number of categories of pedestrian images in all the stored target domain sample subsets. 5.根据权利要求所述3的无监督域自适应行人重识别方法，其特征在于，所述网络模型训练模块，具体通过下述方式获得分类损失函数：5. The unsupervised domain adaptive pedestrian re-identification method according to claim 3, wherein the network model training module obtains a classification loss function in the following manner: 将所述源域样本子集输入行人重识别网络模型的残差网络结构中提取图像特征，以获取源域样本子集中每一行人图片的特征；Inputting the source domain sample subset into the residual network structure of the pedestrian re-identification network model to extract image features to obtain the features of each pedestrian picture in the source domain sample subset; 将该每一行人图片的特征依次输入行人重识别网络模型的全连接层和softmax回归层，进行特征维度转化和特征归一化；The features of each pedestrian image are input into the fully connected layer and softmax regression layer of the pedestrian re-identification network model in turn to perform feature dimension transformation and feature normalization; 基于维度转化和归一化后的每一行人图片的特征采用下述公式分类损失函数；The following formula is used to classify the loss function based on the features of each pedestrian image after dimension transformation and normalization;
Figure FDA0003019905830000032
Figure FDA0003019905830000032
Figure FDA0003019905830000033
Figure FDA0003019905830000033
 其中，Lsrc表示分类损失函数，ns表示源域样本子集中行人图片的数量，xs,m表示源域样本子集中的第m张行人图片，f(xs,m)表示行人图片xs,m的特征，ys,m表示源域样本子集中行人图片xs,m的类别标签，p(ys,m|xs,m)表示行人图片xs,m属于类别ys,m的概率，f(xs,j)表示源域样本子集中第j类行人图片的特征，Ns表示存储的所有源域样本子集中行人图片的类别数。Among them, L src represents the classification loss function, ns represents the number of pedestrian images in the sample subset of the source domain, x s,m represents the mth pedestrian image in the sample subset of the source domain, f(x s,m ) represents the pedestrian image x The features of s, m , y s,m represents the category label of the pedestrian image x s,m in the sample subset of the source domain, p(y s,m | x s,m ) represents the pedestrian image x s,m belongs to the category y s, The probability of m , f(x s,j ) represents the feature of the j-th pedestrian image in the sample subset of the source domain, and N s represents the number of categories of pedestrian images in all the stored sample subsets of the source domain. 6.根据权利要求3所述的无监督域自适应行人重识别方法，其特征在于，所述网络模型训练模块，具体通过下述方式获得样本不变性损失函数：6. The unsupervised domain adaptive pedestrian re-identification method according to claim 3, wherein the network model training module obtains a sample invariance loss function in the following manner: 将所述目标域样本子集输入行人重识别网络模型的残差网络结构中提取图像特征，以获取目标域样本子集中每一行人图片的特征，并将该每一行人图片的特征进行L2标准化处理；Input the target domain sample subset into the residual network structure of the pedestrian re-identification network model to extract image features to obtain the characteristics of each pedestrian picture in the target domain sample subset, and perform L 2 standardized processing; 根据标准化处理后的每一行人图片的特征通过下述公式获得样本不变性损失函数：According to the characteristics of each pedestrian image after normalization, the sample invariance loss function is obtained by the following formula:
Figure FDA0003019905830000041
Figure FDA0003019905830000041
Figure FDA0003019905830000042
Figure FDA0003019905830000042
 其中，LT表示样本不变性损失函数，nt表示目标域样本子集中行人图片的数量，xt,i表示目标域样本子集输入行人重识别网络模型时输入顺序为i的行人图片，f(xt,i)表示行人图片xt,i的特征，wi,l表示行人图片xt,i属于第l类行人图片的权重，p(l|xt,i)表示行人图片xt,i属于第l类行人图片的概率，Nt表示存储的所有目标域样本子集中行人图片的类别数，β为温度系数。Among them, L T represents the sample invariance loss function, n t represents the number of pedestrian images in the target domain sample subset, x t, i represents the pedestrian image in the input order of i when the target domain sample subset is input to the pedestrian re-identification network model, f (x t,i ) represents the feature of the pedestrian image x t,i , w i,l represents the weight of the pedestrian image x t,i belonging to the l-th category of pedestrian images, p(l|x t,i ) represents the pedestrian image x t , i belongs to the probability of the l-th pedestrian image, N t represents the number of categories of pedestrian images in all target domain sample subsets stored, and β is the temperature coefficient. 7.根据权利要求2所述的无监督域自适应行人重识别方法，其特征在于，所述网络模型训练模块，具体通过下述方式对所述行人重识别网络模型进行迭代优化：7. The unsupervised domain self-adaptive pedestrian re-identification method according to claim 2, wherein the network model training module specifically performs iterative optimization on the pedestrian re-identification network model in the following manner: 根据所述分类损失函数、样本不变性损失函数以及分层损失函数获得总损失函数：The total loss function is obtained from the classification loss function, sample invariance loss function and hierarchical loss function: L＝λ1Lsrc+λ2LT+λ3LSL，L=λ 1 L src +λ 2 L T +λ 3 L SL , 其中，L表示总损失函数，Lsrc表示分类损失函数，λ1表示分类损失函数的权重，LT表示样本不变性损失函数，λ2表示样本不变性损失函数的权重，LSL表示分层损失函数，λ3表示分层损失函数的权重；where L represents the total loss function, L src represents the classification loss function, λ 1 represents the weight of the classification loss function, L T represents the sample invariance loss function, λ 2 represents the weight of the sample invariant loss function, and L SL represents the hierarchical loss function, λ 3 represents the weight of the hierarchical loss function; 遍历每一源域样本子集和每一目标域样本子集，迭代更新所述总损失函数，直至总损失函数值的变化量小于预设值，以完成对行人重识别网络模型的优化。Traverse each source domain sample subset and each target domain sample subset, and iteratively update the total loss function until the change in the total loss function value is less than a preset value, so as to complete the optimization of the pedestrian re-identification network model. 8.一种无监督域自适应行人重识别方法，其特征在于，包括：8. An unsupervised domain adaptive pedestrian re-identification method, characterized in that, comprising: 获取包括多个源域样本子集的源域样本集和包括多个目标域样本子集的目标域样本集；obtaining a source domain sample set including multiple source domain sample subsets and a target domain sample set including multiple target domain sample subsets; 基于所述源域样本子集和目标域样本子集分别获得行人重识别网络模型的分类损失函数和样本不变性损失函数，以及根据所述源域样本子集中每一行人图片与目标域样本子集中每一行人图片的相似度对目标域样本子集中的行人图片进行排序、分层，基于分层后的行人图片特征以及对应的层别权重获得分层损失函数；基于所述分类损失函数、样本不变性损失函数以及分层损失函数遍历每一源域样本子集和目标域样本子集对所述行人重识别网络模型进行迭代优化；The classification loss function and the sample invariance loss function of the person re-identification network model are obtained based on the source domain sample subset and the target domain sample subset, respectively, and each pedestrian image in the source domain sample subset and the target domain sample subset are obtained according to Set the similarity of each pedestrian image to sort and layer the pedestrian images in the target domain sample subset, and obtain a layered loss function based on the layered pedestrian image features and the corresponding layer weights; based on the classification loss function, The sample invariance loss function and the hierarchical loss function traverse each source domain sample subset and target domain sample subset to iteratively optimize the pedestrian re-identification network model; 利用优化好的行人重识别网络模型对待识别行人图像进行识别，获得与所述待识别行人图像相同或相似的图像。The optimized pedestrian re-identification network model is used to recognize the image of the pedestrian to be recognized, and an image identical or similar to the image of the pedestrian to be recognized is obtained. 9.根据权利要求8所述的无监督域自适应行人重识别方法，其特征在于，具体通过下述方式获得分层损失函数：9. The unsupervised domain adaptive pedestrian re-identification method according to claim 8, wherein the layered loss function is obtained in the following manner: 将源域样本子集和目标域样本子集分别输入行人重识别网络模型的残差网络结构中提取图像特征，以获取源域样本子集中每一行人图片的特征和目标域样本子集中每一行人图片的特征，并分别进行存储，将源域样本子集中每一行人图片的特征与目标域样本子集中每一行人图片的特征相乘获得对应的相似度；The source domain sample subset and the target domain sample subset are respectively input into the residual network structure of the person re-identification network model to extract image features to obtain the features of each pedestrian image in the source domain sample subset and each row in the target domain sample subset. The features of the person pictures are stored separately, and the features of each pedestrian picture in the source domain sample subset are multiplied by the characteristics of each pedestrian picture in the target domain sample subset to obtain the corresponding similarity; 基于所述相似度将目标域样本子集中的行人图片进行降序排序，并按顺序依次选取第一预设数量的图片作为第一层图片，选取第二预设数量的图片作为第二层图片，其余图片作为第三层图片，并分别设置三层图片的层别权重；Based on the similarity, the pedestrian pictures in the target domain sample subset are sorted in descending order, and the first preset number of pictures are selected as the first layer pictures in sequence, and the second preset number of pictures are selected as the second layer pictures, The rest of the pictures are used as the third-layer pictures, and the layer weights of the three-layer pictures are set respectively; 基于目标域样本子集中每一行人图片的特征以及其所属层别对应的层别权重获得分层损失函数：The hierarchical loss function is obtained based on the characteristics of each pedestrian image in the target domain sample subset and the layer weight corresponding to the layer to which it belongs:
Figure FDA0003019905830000061
Figure FDA0003019905830000061
Figure FDA0003019905830000062
Figure FDA0003019905830000062
 其中，LSL表示分层损失函数，nt表示目标域样本子集中行人图片的数量，wt,r表示层别权重，r表示目标域样本子集中按照相似度排序的行人图片的排序序号，xt,i表示目标域样本子集输入行人重识别网络模型时输入顺序为i的行人图片，f(xt,i)表示行人图片xt,i的特征，p(m|xt,i)表示行人图片xt,i与存储的所有源域样本子集中所有类别的行人图片的相似概率，R1[m]表示存储的所有源域样本子集中所有类别的行人图片的特征，R1[j]表示存储的所有源域样本子集中第j类行人图片的特征，Ns表示存储的所有源域样本子集中行人图片的类别数，β为温度系数。Among them, L SL represents the layered loss function, n t represents the number of pedestrian images in the target domain sample subset, w t, r represents the layer weight, r represents the sequence number of the pedestrian images sorted by similarity in the target domain sample subset, x t,i represents the target domain sample subset inputting the pedestrian image in the order of i when inputting the pedestrian re-identification network model, f(x t, i ) represents the feature of the pedestrian image x t, i , p(m|x t, i ) represents the similarity probability between the pedestrian image x t,i and the stored pedestrian images of all categories in all source domain sample subsets, R 1 [m] represents the stored features of all categories of pedestrian images in all source domain sample subsets, R 1 [j] represents the characteristics of the j-th pedestrian image in all the source domain sample subsets stored, N s represents the number of categories of pedestrian images in all the source domain sample subsets stored, and β is the temperature coefficient. 10.根据权利要求9所述的无监督域自适应行人重识别方法，其特征在于，设置的所述三层图片的层别权重为：10. The unsupervised domain adaptive pedestrian re-identification method according to claim 9, wherein the set layer weights of the three-layer pictures are:
Figure FDA0003019905830000063
Figure FDA0003019905830000063
Figure FDA0003019905830000064
Figure FDA0003019905830000064
 其中，b表示第一层图片的数量，k1表示第一层图片和第二层图片的数量之和，p(l|i)表示行人图片xt,i属于存储的所有目标域样本子集中第l类行人图片的概率，R1[l]表示存储的所有目标域样本子集中第l类行人图片的特征，Nt表示存储的所有目标域样本子集中行人图片的类别数。Among them, b represents the number of pictures in the first layer, k 1 represents the sum of the number of pictures in the first layer and the pictures in the second layer, p(l|i) represents the pedestrian picture x t, i belongs to all the target domain sample subsets stored The probability of the l-th type of pedestrian image, R 1 [l] represents the characteristics of the l-th type of pedestrian image in all the target domain sample subsets stored, and N t represents the number of categories of pedestrian images in all the stored target domain sample subsets.
CN202110399589.7A 2021-04-14 2021-04-14 Self-adaptive pedestrian re-identification system and method for unsupervised domain Active CN113095229B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202110399589.7A CN113095229B (en) 2021-04-14 2021-04-14 Self-adaptive pedestrian re-identification system and method for unsupervised domain

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202110399589.7A CN113095229B (en) 2021-04-14 2021-04-14 Self-adaptive pedestrian re-identification system and method for unsupervised domain

Publications (2)

Publication Number Publication Date
CN113095229A true CN113095229A (en) 2021-07-09
CN113095229B CN113095229B (en) 2024-04-12

Family

ID=76677387

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202110399589.7A Active CN113095229B (en) 2021-04-14 2021-04-14 Self-adaptive pedestrian re-identification system and method for unsupervised domain

Country Status (1)

Country Link
CN (1) CN113095229B (en)

Cited By (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN114792438A (en) * 2022-03-30 2022-07-26 深圳市捷顺科技实业股份有限公司 A face recognition model construction method, device, equipment and storage medium
CN115510926A (en) * 2022-11-23 2022-12-23 武汉理工大学 Diesel engine combustion chamber fault diagnosis method and system across models
CN116405651A (en) * 2023-06-09 2023-07-07 中国矿业大学(北京) Multi-camera cross-view pedestrian data generation method and system
CN119295842A (en) * 2024-12-12 2025-01-10 浙江大学 Generalizable classification method, system and device based on domain-invariant category features

Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20160307072A1 (en) * 2015-04-17 2016-10-20 Nec Laboratories America, Inc. Fine-grained Image Classification by Exploring Bipartite-Graph Labels
CN111126360A (en) * 2019-11-15 2020-05-08 西安电子科技大学 Cross-domain person re-identification method based on unsupervised joint multi-loss model
CN111401281A (en) * 2020-03-23 2020-07-10 山东师范大学 Unsupervised pedestrian re-identification method and system based on deep clustering and sample learning
US20200226421A1 (en) * 2019-01-15 2020-07-16 Naver Corporation Training and using a convolutional neural network for person re-identification

Patent Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20160307072A1 (en) * 2015-04-17 2016-10-20 Nec Laboratories America, Inc. Fine-grained Image Classification by Exploring Bipartite-Graph Labels
US20200226421A1 (en) * 2019-01-15 2020-07-16 Naver Corporation Training and using a convolutional neural network for person re-identification
CN111126360A (en) * 2019-11-15 2020-05-08 西安电子科技大学 Cross-domain person re-identification method based on unsupervised joint multi-loss model
CN111401281A (en) * 2020-03-23 2020-07-10 山东师范大学 Unsupervised pedestrian re-identification method and system based on deep clustering and sample learning

Non-Patent Citations (3)

* Cited by examiner, † Cited by third party
Title
JIAHAN LI等: "Unsupervised Person Re-Identification Based on Measurement Axis", IEEE SIGNAL PROCESSING LETTERS, vol. 28, pages 379 - 383, XP011839499, DOI: 10.1109/LSP.2021.3055116 *
SUNCHENG XIANG等: "Unsupervised person re-identification by hierarchical cluster and domain transfer", MULTIMEDIA TOOLS AND APPLICATIONS, vol. 79, pages 19769 - 19786, XP037214928, DOI: 10.1007/s11042-020-08723-x *
ZHUN ZHONG等: "Invariance Matters: Exemplar Memory for Domain Adaptive Person Re-identification", 2019 IEEE/CVF CONFERENCE ON COMPUTER VISION AND PATTERN RECOGNITION (CVPR), pages 598 - 607 *

Cited By (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN114792438A (en) * 2022-03-30 2022-07-26 深圳市捷顺科技实业股份有限公司 A face recognition model construction method, device, equipment and storage medium
CN115510926A (en) * 2022-11-23 2022-12-23 武汉理工大学 Diesel engine combustion chamber fault diagnosis method and system across models
CN116405651A (en) * 2023-06-09 2023-07-07 中国矿业大学(北京) Multi-camera cross-view pedestrian data generation method and system
CN116405651B (en) * 2023-06-09 2023-08-01 中国矿业大学(北京) Multi-camera cross-view pedestrian data generation method and system
CN119295842A (en) * 2024-12-12 2025-01-10 浙江大学 Generalizable classification method, system and device based on domain-invariant category features
CN119295842B (en) * 2024-12-12 2025-03-21 浙江大学 Generalizable classification method, system and device based on domain-invariant category features

Also Published As

Publication number Publication date
CN113095229B (en) 2024-04-12

Similar Documents

Publication Publication Date Title
CN113095229A (en) Unsupervised domain self-adaptive pedestrian re-identification system and method
CN107194336B (en) Polarization SAR Image Classification Method Based on Semi-supervised Deep Distance Metric Network
CN114612721B (en) Image classification method based on multi-level adaptive feature fusion incremental learning
CN108427740B (en) Image emotion classification and retrieval algorithm based on depth metric learning
CN113326390B (en) Image retrieval method based on deep feature consistent hashing algorithm
CN111310852B (en) An image classification method and system
CN112633382A (en) Mutual-neighbor-based few-sample image classification method and system
CN109492075B (en) Transfer learning sequencing method based on loop generation countermeasure network
CN116089883B (en) A training method used to improve the distinction between old and new categories in incremental learning of existing categories
CN113065516A (en) Unsupervised pedestrian re-identification system and method based on sample separation
CN108647595A (en) Vehicle recognition methods again based on more attribute depth characteristics
CN110941734A (en) Deep unsupervised image retrieval method based on sparse graph structure
CN113139664A (en) Cross-modal transfer learning method
CN111582094A (en) Method for identifying pedestrian by parallel selecting hyper-parameter design multi-branch convolutional neural network
CN109960732A (en) A Robust Supervision-Based Deep Discrete Hash Cross-modal Retrieval Method and System
CN105787045B (en) A kind of precision Enhancement Method for visual media semantic indexing
CN111191033A (en) Open set classification method based on classification utility
CN110796260A (en) A Neural Network Model Optimization Method Based on Class Expansion Learning
CN114298173A (en) A data processing method, device and equipment
CN118628813A (en) Passive domain adaptive image recognition method based on transferable semantic knowledge
CN113283320A (en) Pedestrian re-identification method based on channel feature aggregation
CN114610924A (en) Commodity picture similarity matching search method and system based on multi-layer classification recognition model
CN108304546B (en) A Medical Image Retrieval Method Based on Content Similarity and Softmax Classifier
CN115345248A (en) A data debiasing method and device for deep learning
CN110941994B (en) An ensemble method for person re-identification based on meta-class base learner

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
CB03 Change of inventor or designer information
CB03 Change of inventor or designer information

Inventor after: Cheng Deqiang

Inventor after: Kou Qiqi

Inventor after: Li Jiahan

Inventor after: Li Yunlong

Inventor after: Zhang Haoxiang

Inventor after: Han Chenggong

Inventor after: Xu Jinyang

Inventor after: Jiang Man

Inventor after: Liu Ruihang

Inventor before: Li Jiahan

Inventor before: Li Yunlong

Inventor before: Cheng Deqiang

Inventor before: Kou Qiqi

Inventor before: Zhang Haoxiang

Inventor before: Han Chenggong

Inventor before: Xu Jinyang

Inventor before: Jiang Man

Inventor before: Liu Ruihang

TA01 Transfer of patent application right
TA01 Transfer of patent application right

Effective date of registration: 20240315

Address after: 221116 No. 1 Tongshan University Road, Xuzhou City, Jiangsu Province

Applicant after: CHINA University OF MINING AND TECHNOLOGY

Country or region after: China

Address before: 221116 No. 1 Tongshan University Road, Xuzhou City, Jiangsu Province

Applicant before: CHINA University OF MINING AND TECHNOLOGY

Country or region before: China

Applicant before: Jiangsu Huatu Mining Technology Co.,Ltd.

GR01 Patent grant
GR01 Patent grant