CN112016489B - Pedestrian re-identification method capable of retaining global information and enhancing local features - Google Patents

Pedestrian re-identification method capable of retaining global information and enhancing local features Download PDF

Info

Publication number
CN112016489B
CN112016489B CN202010911071.2A CN202010911071A CN112016489B CN 112016489 B CN112016489 B CN 112016489B CN 202010911071 A CN202010911071 A CN 202010911071A CN 112016489 B CN112016489 B CN 112016489B
Authority
CN
China
Prior art keywords
global
local
branch
pedestrian
information
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.)
Active
Application number
CN202010911071.2A
Other languages
Chinese (zh)
Other versions
CN112016489A (en
Inventor
栾晓
陈俊恒
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.)
Chongqing University of Post and Telecommunications
Original Assignee
Chongqing University of Post and Telecommunications
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 Chongqing University of Post and Telecommunications filed Critical Chongqing University of Post and Telecommunications
Priority to CN202010911071.2A priority Critical patent/CN112016489B/en
Publication of CN112016489A publication Critical patent/CN112016489A/en
Application granted granted Critical
Publication of CN112016489B publication Critical patent/CN112016489B/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
    • G06F18/24—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
    • 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
    • G06F18/2415—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on parametric or probabilistic models, e.g. based on likelihood ratio or false acceptance rate versus a false rejection rate
    • 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
    • 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
    • G06N3/047—Probabilistic or stochastic networks
    • 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
    • G06N3/08—Learning methods

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Data Mining & Analysis (AREA)
  • General Physics & Mathematics (AREA)
  • Evolutionary Computation (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Artificial Intelligence (AREA)
  • General Engineering & Computer Science (AREA)
  • Computing Systems (AREA)
  • General Health & Medical Sciences (AREA)
  • Molecular Biology (AREA)
  • Computational Linguistics (AREA)
  • Biophysics (AREA)
  • Biomedical Technology (AREA)
  • Mathematical Physics (AREA)
  • Software Systems (AREA)
  • Health & Medical Sciences (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Evolutionary Biology (AREA)
  • Probability & Statistics with Applications (AREA)
  • Human Computer Interaction (AREA)
  • Multimedia (AREA)
  • Image Analysis (AREA)
  • Image Processing (AREA)

Abstract

The invention discloses a pedestrian re-identification method for retaining global information and enhancing local features, which comprises the following steps of: s1: changing the size of the original pedestrian image to 384 multiplied by 128 multiplied by 3; s2: respectively extracting global characteristic information and local characteristic information of the pedestrian through global-branch and local-branch; s3: the fusion guide module fuses the global feature information with each local feature information respectively: s4: inputting the features after global average pooling of global-branch and local-branch into a triple loss function for metric learning, and inputting the features of the global-branch, local-branch and fusion guide module into a cross entropy loss function for classification learning; s5: inputting the characteristics of the global-branch, the local-branch and the fusion guide module into respective trained classifiers, and outputting the result of pedestrian re-identification classification. The invention can weaken the problems of shielding, image blurring and pedestrian misalignment in the pedestrian image.

Description

Pedestrian re-identification method capable of retaining global information and enhancing local features
Technical Field
The invention relates to the field of digital image processing and pattern recognition, in particular to a pedestrian re-recognition method for retaining global information and enhancing local features.
Background
Pedestrian re-identification refers to a technology for judging whether a specific pedestrian exists in an image or a video sequence by using a computer vision technology. Given a monitored pedestrian image, the pedestrian image is retrieved across the device. The technology aims to make up the visual limitation of a fixed camera, can be combined with a pedestrian detection/pedestrian tracking technology, and can be widely applied to the fields of intelligent video monitoring, intelligent security and the like. Because the cost of manual data labeling is high, most data samples are used for pedestrian image under the monitoring video selected by the pedestrian detector, and the problems of shielding, image blurring, pedestrian misalignment and the like generally exist. In addition, the existing method based on local features ignores or does not fully utilize the global semantic information of pedestrians, only focuses on mining local information, and does not combine and utilize the advantages of the global semantic information and the local semantic information.
Disclosure of Invention
The invention provides a pedestrian re-identification method for retaining global information and enhancing local features, and aims to solve the problems of shielding, image blurring and pedestrian misalignment in a pedestrian image.
The invention is realized by the following technical scheme:
a pedestrian re-identification method for retaining global information and enhancing local characteristics is applied to a corresponding network model and comprises the following steps:
s1, a data preparation stage: for the pedestrian image, the size of the original pedestrian image is changed into 384 multiplied by 128 multiplied by 3 in the network model, wherein the 384 multiplied by 128 multiplied by 3 corresponds to the height, the width and the channel number of the image;
s2, a characteristic extraction stage: global feature information and local feature information of pedestrians are respectively extracted through global-branch and local-branch, and then the obtained information is subjected to global average pooling and global maximum pooling, wherein: the global characteristic refers to the overall attribute of the image target, the local characteristic refers to the local attribute of the image target, and in the network model, the network branch responsible for extracting the global characteristic information is named as a global branch, and the network branch responsible for extracting the local characteristic information is named as a local branch;
s3, feature fusion and guidance stage: the fusion guide module in the network model fuses the global feature information with each local feature information respectively, and the fused feature information not only retains the global semantic information of the pedestrian, but also enhances the expression capability of the corresponding local features fused with the global feature information; with the reverse conduction and gradient updating of the network model training process, the fusion guide module further improves the global-branch and local-branch feature extraction capability;
s4, model training stage: inputting the characteristic information after the pooling of the global-branch and the local-branch into a triple loss function for metric learning, and inputting the characteristics of the global-branch, the local-branch and the fusion guide module into a cross entropy loss function for classification learning;
s5, model evaluation stage: inputting the characteristics of the global-branch, the local-branch and the fusion guide module into respective trained classifiers, and outputting the result of pedestrian re-identification classification.
Further, S1 specifically is:
the original pedestrian images i of the data set in the network model are uniform in size, and the conversion formula is as follows:
I＝resize(i)
wherein the size of I is 384 × 128 × 3.
Further, in S2:
inputting a pedestrian image I into a network model, then adopting ResNet-50 as a main network to carry out primary feature extraction, and taking a feature map output by a 3 rd residual block of ResNet-50 as T b3 The size is 24 × 8 × 1024.
Further, in S2:
in global-branch, T is b3 Inputting the original 4 th residual block of ResNet-50 to obtain a characteristic diagram T b4 The size is 12 × 4 × 2048; here, the 1 × 1 convolution, 3 × 3 convolution and upsampling operations are denoted as Conv1, conv3 and Upsample, respectively; thus, the following formula is passed:
T g ＝Conv1(Conv3(T b3 +Conv1(Upsample(T b4 ))))
the global information enhancement module can obtain the feature map T g The size is 24 × 8 × 2048; obtaining global characteristics after global average pooling and global maximum pooling
Figure BDA0002663279520000021
The size of the drug is 1 multiplied by 2048; the pooling formula is as follows:
Figure BDA0002663279520000022
further, in S2:
in local-branching, the downsampling operation of the 4 th residual block of ResNet-50 is removed, T b3 Inputting the residual block to obtain a feature map T p The size of which is 24 × 8 × 2048; obtaining local characteristics after global average pooling and global maximum pooling
Figure BDA0002663279520000023
The size of the drug is 1 multiplied by 2048; the pooling equation is as follows:
Figure BDA0002663279520000024
further, in S3:
in the fusion guidance module, global features
Figure BDA0002663279520000025
With each local feature
Figure BDA0002663279520000026
Carry out the para-position addition to obtain the fusion characteristic
Figure BDA0002663279520000027
The formula is as follows:
Figure BDA0002663279520000031
further, in S4:
will the characteristic diagram T g And the feature map T p After global average pooling, respectively obtaining
Figure BDA0002663279520000032
And
Figure BDA0002663279520000033
the size is 1 × 1 × 2048, and the pooling formula is as follows:
Figure BDA0002663279520000034
Figure BDA0002663279520000035
further, in S4:
will be provided with
Figure BDA0002663279520000036
And
Figure BDA0002663279520000037
feature metric learning is carried out in input loss 1, the loss 1 is calculated by adopting a triple loss function, and the formula is as follows:
Figure BDA0002663279520000038
wherein N represents the total number of input Triplet Loss samples in the network model,
Figure BDA0002663279520000039
and
Figure BDA00026632795200000310
respectively representing the Anchor, positive and Negative samples in the network model, alpha is a difference value hyper-parameter [ ·] + ＝max(·,0)。
Further, in S4:
global feature vector
Figure BDA00026632795200000311
Local feature vector
Figure BDA00026632795200000312
And fusion feature vectors
Figure BDA00026632795200000313
Respectively inputting full-connection layers for dimensionality reduction, wherein the dimensionality reduction size is 1 multiplied by 256, then inputting the loss 2 for classification training, wherein the loss 2 is calculated by adopting a cross entropy loss function, and the formula is as follows:
Figure BDA00026632795200000314
wherein, W k Representing weight vectors of k classes, N representing the total number of input Softmax Loss samples in the network model, and C representing the total number of sample classes; and after the network model training is finished, obtaining a trained classifier corresponding to the feature vector.
Further, in S5:
and respectively inputting the feature vectors of which the dimensions are reduced to 1 multiplied by 256 into corresponding trained classifiers, and outputting the classification result of the re-identification of the pedestrians.
Compared with the prior art, the invention has the following advantages and beneficial effects:
1. the invention designs a global information enhancement module. The residual error neural network ResNet-50 comprises 4 residual error blocks, and the invention can extract richer global feature information of pedestrian semantics by fusing feature maps output by 2 residual error blocks with less memory and parameter cost.
2. The invention designs a fusion guide module which can fuse two types of characteristics of global-branch and local-branch, and enhances the expression capability of corresponding local characteristics fused with the global semantic information of pedestrians while keeping the global semantic information of the pedestrians. And with the reverse conduction of the training process, the fusion guide module further improves the global-branch and local-branch feature extraction capability.
Drawings
The accompanying drawings, which are included to provide a further understanding of the embodiments of the invention and are incorporated in and constitute a part of this application, illustrate embodiment(s) of the invention and together with the description serve to explain the principles of the invention. In the drawings:
fig. 1 is a schematic diagram of a network model structure according to the present invention.
Detailed Description
In order to make the objects, technical solutions and advantages of the present invention more apparent, the present invention is further described in detail below with reference to examples and accompanying drawings, and the exemplary embodiments and descriptions thereof are only used for explaining the present invention and are not meant to limit the present invention.
Example (b):
s1: a data preparation stage: the size of the original pedestrian image is changed to 384 × 128 × 3 (corresponding to the height, width and number of channels of the image, respectively):
for the pedestrian original image i of the data set, the size is unified, and the conversion formula is as follows:
I＝resize(i) (1)
where the size of I is 384 × 128 × 3 (image height × image width × number of image channels).
S2: a characteristic extraction stage: respectively extracting global characteristic information and local characteristic information of the pedestrian through global-branch and local-branch;
s21, inputting a pedestrian image I into a network model, then adopting ResNet-50 as a main network to carry out primary feature extraction, and taking a feature graph output by a 3 rd residual block of ResNet-50 as T b3 The size of which is 24 × 8 × 1024;
s22: in global-branch, T of S21 b3 Inputting the original 4 th residual block of ResNet-50 to obtain a characteristic diagram T b4 The size is 12 × 4 × 2048. Here, the 1 × 1 convolution, the 3 × 3 convolution and the upsampling operation are denoted as Conv1, conv3 and Upsample, respectively. Thus, the following formula is passed:
T g ＝Conv1(Conv3(T b3 +Conv1(Upsample(T b4 )))) (2)
the global information enhancement module can obtain the feature map T g The size is 24 × 8 × 2048. After Global Average Pooling (GAP) and Global Maximum Pooling (GMP), global characteristics are obtained
Figure BDA0002663279520000041
The size is 1 × 1 × 2048. Pooling formulaThe following were used:
Figure BDA0002663279520000051
s23: in local-branching, the downsampling operation of the 4 th residual block of ResNet-50 is deleted, T of S21 b3 Inputting the residual block to obtain a feature map T p The size is 24 × 8 × 2048. After GAP and GMP, local characteristics are obtained
Figure BDA0002663279520000052
The size is 1 × 1 × 2048. The pooling equation is as follows:
Figure BDA0002663279520000053
s3: and (3) feature fusion and guide stage: the fusion guide module fuses the global feature information with each local feature information respectively, and the fused features not only keep the global semantic information of the pedestrians, but also enhance the expression capability of the corresponding local features fused with the global feature information. Besides, with the reverse conduction of the training process, the fusion guide module further improves the global-branch and local-branch feature extraction capability:
s31: in the fusion guidance module, the global feature of S22
Figure BDA0002663279520000054
And each local feature of S23
Figure BDA0002663279520000055
Carry out the para-position addition to obtain the fusion characteristic
Figure BDA0002663279520000056
The formula is as follows:
Figure BDA0002663279520000057
and S32, conducting reversely and updating the gradient in the training process, and fusing the guide module to further improve the global-branch and local-branch feature extraction capability.
S4, model training: inputting the feature information after the pooling of the global-branch and the local-branch into a triple loss function for metric learning, and inputting the features of the global-branch, the local-branch and the fusion guide module into a cross entropy loss function for classification learning:
s41: the characteristic diagram T of S22 g And comparing the characteristic map T of S23 p After GAP, respectively obtain
Figure BDA0002663279520000058
And
Figure BDA0002663279520000059
the size is 1 × 1 × 2048. The pooling equation is as follows:
Figure BDA00026632795200000510
Figure BDA00026632795200000511
s42: will be provided with
Figure BDA00026632795200000512
And
Figure BDA00026632795200000513
the feature metric learning is performed in the loss 1 shown in fig. 1. In the invention, the Loss 1 adopts a triple Loss function (triple Loss), and the formula is as follows:
Figure BDA00026632795200000514
wherein N represents the total number of input triple Loss samples,
Figure BDA00026632795200000515
and
Figure BDA00026632795200000516
respectively representing an Anchor, positive and Negative sample, alpha being a difference hyperparameter [ · C] + ＝max(·,0)。
S43: global feature vector of S22
Figure BDA0002663279520000061
S23 local feature vector
Figure BDA0002663279520000062
And the fused feature vector of S31
Figure BDA0002663279520000063
Respectively inputting the full-connection layers for dimension reduction, wherein the dimension reduction is 1 multiplied by 256, and then inputting the dimension reduction into a loss 2 for classification training. In the invention, the Loss 2 adopts a cross entropy Loss function (Softmax Loss), and the formula is as follows:
Figure BDA0002663279520000064
wherein, W k The weight vector representing class k, N the total number of input Softmax Loss samples, and C the total number of sample classes. And after the network model training is finished, obtaining a trained classifier corresponding to the feature vector.
S5: in the model evaluation stage, the feature vectors with dimension reduced from S43 to 1 × 1 × 256 are respectively input into the trained classifier corresponding to S43, and the classification result of pedestrian re-identification is output.
Based on a pedestrian re-identification method which reserves global information and enhances local features, a network model structure shown in figure 1 is constructed, and tests are carried out on a Market-1501 and DukeMTMC-ReiD data set. In the field of pedestrian re-identification research, a Cumulative Matching Curve (CMC) and an Average Precision Average (mAP) are generally used to evaluate the performance of a method on a pedestrian re-identification data set. Meanwhile, in table 1, a method 1 based on local features, which employs horizontal blocking of a feature map, a method 2 employing a two-branch network structure, a method 3 employing a three-branch network, which gradually transits from global feature learning to local feature learning, and a method 4 employing the present invention are compared.
The following table gives the results of the tests on the database, and it can be seen that the network model based on the present invention performed more excellently on each data set result, both in the CMC (Rank-1 and Rank 5) and in the mAP indices.
Figure BDA0002663279520000065
The above-mentioned embodiments, objects, technical solutions and advantages of the present invention are further described in detail, it should be understood that the above-mentioned embodiments are only examples of the present invention, and are not intended to limit the scope of the present invention, and any modifications, equivalent substitutions, improvements and the like made within the spirit and principle of the present invention should be included in the scope of the present invention.

Claims (9)

1. A pedestrian re-identification method for retaining global information and enhancing local features, which is applied in a corresponding network model, is characterized by comprising the following steps:
s1, a data preparation stage: for the pedestrian image, the size of the original pedestrian image is changed into 384 multiplied by 128 multiplied by 3 in the network model, wherein the 384 multiplied by 128 multiplied by 3 corresponds to the height, the width and the channel number of the image;
s2, a characteristic extraction stage: global feature information and local feature information of pedestrians are respectively extracted through global-branch and local-branch, and then the obtained information is subjected to global average pooling and global maximum pooling, wherein: the global characteristic refers to the overall attribute of the image target, the local characteristic refers to the local attribute of the image target, and in the network model, the network branch responsible for extracting the global characteristic information is named as a global branch, and the network branch responsible for extracting the local characteristic information is named as a local branch;
s3, a characteristic fusion and guide stage: the fusion guide module in the network model fuses the global feature information with each local feature information respectively, and the fused feature information not only retains the global semantic information of the pedestrian, but also enhances the expression capability of the corresponding local features fused with the global feature information; with the reverse conduction and gradient updating in the network model training process, the fusion guide module further improves the global-branch and local-branch feature extraction capability;
s4, model training stage: inputting the pooled feature information of the global-branch and the local-branch into a triple loss function for metric learning, and inputting the features of the global-branch, the local-branch and the fusion guide module into a cross entropy loss function for classification learning;
s5, model evaluation stage: inputting the global feature vector, the local feature vector and the fusion feature vector into a full-connection layer respectively for dimensionality reduction, and then inputting the global feature vector, the local feature vector and the fusion feature vector into a cross entropy loss function for classification training; after the network model training is finished, obtaining a trained classifier corresponding to the feature vector;
and respectively inputting the feature vectors of which the dimensions are reduced to 1 multiplied by 256 into the trained classifier, and outputting the classification result of the re-identification of the pedestrians.
2. The pedestrian re-identification method for retaining global information and enhancing local features according to claim 1, wherein S1 specifically is:
the original pedestrian images i of the data set in the network model are uniform in size, and the conversion formula is as follows:
I＝resize(i)
wherein the size of I is 384 × 128 × 3.
3. The pedestrian re-identification method for retaining global information and enhancing local features as claimed in claim 2, wherein in S2:
inputting the pedestrian image I into a network model, then adopting ResNet-50 as a main network to carry out primary feature extraction, and taking ResNet-50The feature map output by the 3 rd residual block is T b3 The size is 24 × 8 × 1024.
4. The pedestrian re-identification method for retaining global information and enhancing local features as claimed in claim 3, wherein in S2:
in global-branch, T is b3 Inputting the original 4 th residual block of ResNet-50 to obtain a characteristic diagram T b4 The size is 12 × 4 × 2048; here, the 1 × 1 convolution, the 3 × 3 convolution and the upsampling operation are denoted as Conv1, conv3 and Upsample, respectively; thus, the following formula is passed:
T g ＝Conv1(Conv3(T b3 +Conv1(Upsample(T b4 ))))
the global information enhancement module can obtain the feature map T g The size is 24 × 8 × 2048; obtaining global characteristics after global average pooling and global maximum pooling
Figure FDA0003796092800000021
The size is 1 × 1 × 2048; the pooling equation is as follows:
Figure FDA0003796092800000022
5. the pedestrian re-identification method for retaining global information and enhancing local features as claimed in claim 4, wherein in S2:
in local-branching, the downsampling operation of the 4 th residual block of ResNet-50 is removed, T b3 Inputting the residual block to obtain a feature map T p The size of which is 24 × 8 × 2048; obtaining local characteristics after global average pooling and global maximum pooling
Figure FDA0003796092800000023
The size of the drug is 1 multiplied by 2048; the pooling formula is as follows:
Figure FDA0003796092800000024
6. the pedestrian re-identification method for retaining global information and enhancing local features as claimed in claim 5, wherein in S3:
in the fusion guidance module, global features
Figure FDA0003796092800000025
And each local feature
Figure FDA0003796092800000026
Carry out the para-position addition to obtain the fusion characteristic
Figure FDA0003796092800000027
The formula is as follows:
Figure FDA0003796092800000028
7. the pedestrian re-identification method for retaining global information and enhancing local features as claimed in claim 6, wherein in S4:
feature map T g And the feature map T p Respectively obtain after global average pooling
Figure FDA0003796092800000029
And
Figure FDA00037960928000000210
the size is 1 × 1 × 2048, and the pooling formula is as follows:
Figure FDA00037960928000000211
Figure FDA00037960928000000212
8. the pedestrian re-identification method for retaining global information and enhancing local features as claimed in claim 7, wherein in S4:
will be provided with
Figure FDA0003796092800000031
And
Figure FDA0003796092800000032
feature metric learning is carried out in input loss 1, the loss 1 is calculated by adopting a triple loss function, and the formula is as follows:
Figure FDA0003796092800000033
wherein N represents the total number of input Triplet Loss samples in the network model,
Figure FDA0003796092800000034
and
Figure FDA0003796092800000035
respectively representing the Anchor, positive and Negative samples in the network model, alpha is a difference value hyper-parameter [ ·] + ＝max(·,0)。
9. The pedestrian re-identification method for retaining global information and enhancing local features as claimed in claim 8, wherein in S4:
global feature vector
Figure FDA0003796092800000036
Local feature vector
Figure FDA0003796092800000037
And fusion feature vectors
Figure FDA0003796092800000038
Respectively inputting full-connection layers for dimensionality reduction, wherein the dimensionality reduction size is 1 multiplied by 256, and then inputting cross entropy loss function calculation, wherein the formula is as follows:
Figure FDA0003796092800000039
wherein, W k Representing weight vectors of k classes, N representing the total number of input Softmax Loss samples in the network model, and C representing the total number of sample classes; and after the network model training is finished, obtaining a trained classifier corresponding to the feature vector.
CN202010911071.2A 2020-09-02 2020-09-02 Pedestrian re-identification method capable of retaining global information and enhancing local features Active CN112016489B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202010911071.2A CN112016489B (en) 2020-09-02 2020-09-02 Pedestrian re-identification method capable of retaining global information and enhancing local features

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202010911071.2A CN112016489B (en) 2020-09-02 2020-09-02 Pedestrian re-identification method capable of retaining global information and enhancing local features

Publications (2)

Publication Number Publication Date
CN112016489A CN112016489A (en) 2020-12-01
CN112016489B true CN112016489B (en) 2022-10-04

Family

ID=73516709

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202010911071.2A Active CN112016489B (en) 2020-09-02 2020-09-02 Pedestrian re-identification method capable of retaining global information and enhancing local features

Country Status (1)

Country Link
CN (1) CN112016489B (en)

Families Citing this family (7)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112560604A (en) * 2020-12-04 2021-03-26 中南大学 Pedestrian re-identification method based on local feature relationship fusion
CN112560932A (en) * 2020-12-10 2021-03-26 山东建筑大学 Vehicle weight identification method based on dual-branch network feature fusion
CN112801235A (en) * 2021-04-12 2021-05-14 四川大学 Model training method, prediction device, re-recognition model and electronic equipment
CN112927171A (en) * 2021-04-15 2021-06-08 重庆邮电大学 Single image deblurring method based on generation countermeasure network
CN113177464B (en) * 2021-04-27 2023-12-01 浙江工商大学 End-to-end multi-mode gait recognition method based on deep learning
CN113239784B (en) * 2021-05-11 2022-09-30 广西科学院 Pedestrian re-identification system and method based on space sequence feature learning
CN114550315A (en) * 2022-01-24 2022-05-27 云南联合视觉科技有限公司 Identity comparison and identification method and device and terminal equipment

Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN108960140A (en) * 2018-07-04 2018-12-07 国家新闻出版广电总局广播科学研究院 The pedestrian's recognition methods again extracted and merged based on multi-region feature
CN109784258A (en) * 2019-01-08 2019-05-21 华南理工大学 A kind of pedestrian's recognition methods again cut and merged based on Analysis On Multi-scale Features
WO2020125216A1 (en) * 2018-12-18 2020-06-25 深圳云天励飞技术有限公司 Pedestrian re-identification method, device, electronic device and computer-readable storage medium
CN111539370A (en) * 2020-04-30 2020-08-14 华中科技大学 Image pedestrian re-identification method and system based on multi-attention joint learning

Family Cites Families (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109191558B (en) * 2018-07-27 2020-12-08 深圳市商汤科技有限公司 Image polishing method and device

Patent Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN108960140A (en) * 2018-07-04 2018-12-07 国家新闻出版广电总局广播科学研究院 The pedestrian's recognition methods again extracted and merged based on multi-region feature
WO2020125216A1 (en) * 2018-12-18 2020-06-25 深圳云天励飞技术有限公司 Pedestrian re-identification method, device, electronic device and computer-readable storage medium
CN109784258A (en) * 2019-01-08 2019-05-21 华南理工大学 A kind of pedestrian's recognition methods again cut and merged based on Analysis On Multi-scale Features
CN111539370A (en) * 2020-04-30 2020-08-14 华中科技大学 Image pedestrian re-identification method and system based on multi-attention joint learning

Non-Patent Citations (3)

* Cited by examiner, † Cited by third party
Title
multi-scale and multi-branch feature representation for person re-identification;shanshan jiao 等;《Neurocomputing》;20200630;120-130 *
基于深度学习的行人重识别研究;陈俊恒;《中国优秀硕士学位论文全文数据库信息科技辑》;20220315(第03期);I138-2513 *
多粒度特征融合的行人再识别研究;张良等;《液晶与显示》;20200630;第35卷(第06期);46-54 *

Also Published As

Publication number Publication date
CN112016489A (en) 2020-12-01

Similar Documents

Publication Publication Date Title
CN112016489B (en) Pedestrian re-identification method capable of retaining global information and enhancing local features
CN110084124B (en) Feature enhancement target detection method based on feature pyramid network
US20220230282A1 (en) Image processing method, image processing apparatus, electronic device and computer-readable storage medium
CN107844743B (en) Image multi-subtitle automatic generation method based on multi-scale hierarchical residual error network
CN113569865B (en) Single sample image segmentation method based on class prototype learning
CN114119993B (en) Remarkable target detection method based on self-attention mechanism
CN115082675B (en) Transparent object image segmentation method and system
CN111652081A (en) Video semantic segmentation method based on optical flow feature fusion
US12056950B2 (en) Transformer-based multi-scale pedestrian re-identification method
CN116343043B (en) Remote sensing image change detection method with multi-scale feature fusion function
Yuan et al. Low altitude small UAV detection based on YOLO model
CN116128998A (en) Multi-path parallel text-to-image generation method and system
CN117197687A (en) Unmanned aerial vehicle aerial photography-oriented detection method for dense small targets
CN111079543A (en) Efficient vehicle color identification method based on deep learning
Wang et al. Pedestrian detection based on spatial attention module for outdoor video surveillance
Liu et al. Semantic segmentation of high-resolution remote sensing images using an improved transformer
Li et al. Improved YOLOv5s algorithm for small target detection in UAV aerial photography
CN117391962A (en) Classification task-oriented remote sensing image super-resolution method and system
CN112418229A (en) Unmanned ship marine scene image real-time segmentation method based on deep learning
CN111753714A (en) Multidirectional natural scene text detection method based on character segmentation
Shen et al. StepNet: Spatial-temporal part-aware network for sign language recognition
CN111767919A (en) Target detection method for multi-layer bidirectional feature extraction and fusion
CN114972760B (en) Ionization diagram automatic tracing method based on multi-scale attention-enhancing U-Net
CN116152199A (en) Hand gesture and shape estimation method based on segmentation map guidance and regular constraint
CN116342875A (en) Small sample segmentation method based on multi-scale self-attention feature 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
GR01 Patent grant
GR01 Patent grant