CN113449552A - Pedestrian re-identification method based on blocking indirect coupling GAN network - Google Patents
Pedestrian re-identification method based on blocking indirect coupling GAN network Download PDFInfo
- Publication number
- CN113449552A CN113449552A CN202010218063.XA CN202010218063A CN113449552A CN 113449552 A CN113449552 A CN 113449552A CN 202010218063 A CN202010218063 A CN 202010218063A CN 113449552 A CN113449552 A CN 113449552A
- Authority
- CN
- China
- Prior art keywords
- network
- pedestrian
- gan network
- layer
- feature vector
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
- 238000000034 method Methods 0.000 title claims abstract description 31
- 230000008878 coupling Effects 0.000 title claims abstract description 23
- 238000010168 coupling process Methods 0.000 title claims abstract description 23
- 238000005859 coupling reaction Methods 0.000 title claims abstract description 23
- 230000000903 blocking effect Effects 0.000 title claims abstract description 17
- 238000012549 training Methods 0.000 claims abstract description 15
- 238000012512 characterization method Methods 0.000 claims abstract description 4
- 239000013598 vector Substances 0.000 claims description 66
- 238000005070 sampling Methods 0.000 claims description 12
- 230000006870 function Effects 0.000 claims description 7
- 230000008569 process Effects 0.000 claims description 7
- 239000011159 matrix material Substances 0.000 claims description 5
- 238000000605 extraction Methods 0.000 claims description 4
- 230000004927 fusion Effects 0.000 claims description 4
- 238000011176 pooling Methods 0.000 claims description 4
- 101150041570 TOP1 gene Proteins 0.000 claims description 3
- 230000001174 ascending effect Effects 0.000 claims description 3
- 238000010606 normalization Methods 0.000 claims description 3
- 238000012545 processing Methods 0.000 claims description 3
- 238000012163 sequencing technique Methods 0.000 claims description 2
- 230000036544 posture Effects 0.000 abstract description 16
- 230000008859 change Effects 0.000 abstract description 7
- 238000005286 illumination Methods 0.000 abstract description 2
- 238000010586 diagram Methods 0.000 description 4
- 230000037237 body shape Effects 0.000 description 2
- 238000005034 decoration Methods 0.000 description 2
- 238000013135 deep learning Methods 0.000 description 2
- 238000012986 modification Methods 0.000 description 2
- 230000004048 modification Effects 0.000 description 2
- 230000000295 complement effect Effects 0.000 description 1
- 230000007547 defect Effects 0.000 description 1
- 238000011161 development Methods 0.000 description 1
- 238000005516 engineering process Methods 0.000 description 1
- 238000012886 linear function Methods 0.000 description 1
- 238000005259 measurement Methods 0.000 description 1
- 238000011524 similarity measure Methods 0.000 description 1
- 230000000007 visual effect Effects 0.000 description 1
Images
Classifications
-
- 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
- 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
- G06N3/08—Learning methods
- G06N3/084—Backpropagation, e.g. using gradient descent
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- Evolutionary Computation (AREA)
- Life Sciences & Earth Sciences (AREA)
- Artificial Intelligence (AREA)
- General Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- General Health & Medical Sciences (AREA)
- Software Systems (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- Biophysics (AREA)
- Biomedical Technology (AREA)
- Mathematical Physics (AREA)
- Computational Linguistics (AREA)
- Health & Medical Sciences (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Evolutionary Biology (AREA)
- Image Analysis (AREA)
Abstract
The invention discloses a pedestrian re-identification method based on a blocking indirect coupling GAN network. And then, the partitioned indirect coupling characterization learning network is used as an encoder of the GAN network, namely the partitioned indirect coupling GAN network is used for training the GAN network, and the same pedestrian features capable of being recognized under different postures are obtained through back propagation of a discriminator, a decoder and the like. Therefore, more robust features are obtained, and pedestrian re-identification is carried out. When the pedestrian re-identification is finally carried out, only an encoder in the GAN network is used for extracting features, and extra attitude information and computing power are not needed; for the illumination change, the recognition rate of different postures of pedestrians with the same scene change and posture change is higher; different pedestrians have strong distinguishing force due to clothes similarity and the like; the generated graph is clearer and contains more detailed information.
Description
Technical Field
The invention belongs to the technical field of pedestrian re-identification, and particularly relates to a pedestrian re-identification method based on a blocking indirect coupling GAN network.
Background
Pedestrian Re-identification (Re-ID) is an important application in security and intelligent surveillance systems, and is an effective alternative when face recognition cannot be used. The Re-ID is intended to search for the same pedestrian under different cameras according to the characteristics of the body shape, appearance and posture of the person. Initially, the pedestrian Re-ID method was primarily based on traditional algorithms, including manual extraction of visual features and similarity measures. Different from the traditional method, the deep learning method can automatically extract better features and learn better similarity indexes. With the development of deep learning, the pedestrian recognition technology has also made great progress. However, pedestrians in real scenes are often occluded by moving objects or stationary objects. And the pedestrian to camera distance is also not fixed, which will result in low resolution objects and large scale variations. In addition, the same person may have large deformation and posture change under different cameras. Different pedestrians may have very similar dresses, poses, body shapes and appearances. Therefore, pedestrian re-identification remains a hot problem in recent years.
Block-based representation learning methods have proven to be very effective methods for pedestrian Re-identification (Re-ID) and have a fast convergence rate, but the features extracted by existing block-based methods tend to have a high correlation between different blocks. And the Re-ID method based on expression learning is less effective for pedestrians with large posture variations. In order to reduce the influence of posture difference and body obstruction, a generation countermeasure network (GAN) is applied in the Re-ID field, which mainly requires that extracted features exist in different posture images when another posture is generated according to the features of an input image in the training process, so that the extracted features can better identify the same pedestrian in different postures, but the features extracted by the current GAN network-based Re-ID method have the problem of redundancy for a small part, so that the generated map contains less detailed information and is fuzzy.
Disclosure of Invention
The purpose of the invention is as follows: aiming at the defects in the prior art, the invention aims to provide a blocking indirect coupling GAN network-based method. And then, the partitioned indirect coupling characterization learning network is used as an encoder of the GAN network, namely the partitioned indirect coupling GAN network is used for training the GAN network, and the same pedestrian features capable of being recognized under different postures are obtained through back propagation of a discriminator, a decoder and the like. Therefore, more robust features are obtained, and pedestrian re-identification is carried out.
The technical scheme is as follows: in order to achieve the purpose of the invention, the technical scheme adopted by the invention is as follows: the pedestrian re-identification method based on the blocking indirect coupling GAN network comprises the following steps:
step 1: inputting a pedestrian image into an encoder in the GAN network; the GAN network consists of an encoder, a decoder and two discriminators; the two discriminators are an identity discriminator Did and a gesture discriminator Dpd respectively;
step 2: training an encoder to obtain characteristics capable of distinguishing different pedestrians;
and step 3: inputting the pedestrian picture again, and extracting the pedestrian features through an encoder; inputting a target attitude heat map, and obtaining attitude characteristics through an attitude characteristic extraction network; then combining the pedestrian feature vector, the attitude feature vector and the noise vector together in a channel dimension through concatee operation, inputting the combined pedestrian feature vector, attitude feature vector and noise vector into a decoder in a GAN network, and then training the combined pedestrian feature vector, attitude feature vector and noise vector to obtain a generated image;
and 4, step 4: an identity discriminator D for inputting the input pedestrian image, the target posture original image and the generated image into the GAN networkidPerforming the following steps;
and 5: an attitude discriminator D for inputting the input pedestrian image attitude map, the target attitude map and the generated map into the GAN networkpdPerforming the following steps;
step 6: training the whole GAN network, extracting the characteristics of the query graph and the graph to be queried by using the coding network in the GAN network, and calculating the similarity of the characteristics for re-recognition.
Further, the Encoder in step 2 is composed of a partitioned indirectly coupled characterization learning network (i.e. an encoding network Encoder), and step 2 includes:
step 2-1: uniformly scaling the pedestrian images to 384 × 128, and performing image enhancement operations such as horizontal turning, random erasing and the like on the images;
step 2-2: inputting the graph obtained in the step 2-1 into a coding network for training, horizontally dividing the obtained feature vectors into 6 blocks, obtaining 256-dimensional column vectors for each block, and calculating cross entropy loss with a real class label;
step 2-3: the vector of each image block obtained in the step 2-1 is calculated, the cosine similarity of any two characteristic vectors is calculated, and L1 loss is measured together with the zero vector;
further, the encoding network Encoder is composed of a reference network and two loss functions, and is specifically described as follows:
for step 2-2, for each of the 6 blocks, performing an average pooling and 1x1 convolution operation; obtaining a 256-dimensional column vector for each block;
and 2-3, performing L2 regularization on the network weight matrix when the cosine similarity is close to zero, thereby obtaining the characteristic of segment independence.
Further, the decoding network Decoder of the Decoder in step 3 is an up-sampling process, and step 3 includes:
step 3-1: inputting the fused features, and obtaining a feature vector of a layer 1 through the operation of a deconvolution module;
step 3-2: inputting the output of the 1 st layer into a deconvolution module to obtain a feature vector of the 2 nd layer;
step 3-3: the up-sampling operation of the feature vector of the layer 2 integrates a deconvolution module and a bilinear interpolation module, so that a layer 3 feature vector is obtained;
step 3-4: the feature vector of the layer 3 is subjected to the up-sampling operation of the fusion module to obtain a feature vector of the layer 4;
step 3-5: and the feature vector of the 4 th layer is subjected to a deconvolution module to obtain a generated graph.
Further, the acquisition process on the decoding network Decoder is as follows: the network input layer fuses the pedestrian characteristics, the target attitude characteristics and the combined characteristics of the noise vectors, then carries out up-sampling operation and decodes the combined characteristics into the size of an input graph;
the sub-modules sampled by the decoder are composed of a deconvolution module and a Bilinear interpolation module, the Bilinear interpolation module is composed of a Biliner-Cov network, and a 1x1 convolution layer is added to obtain the same number of characteristic channels as the deconvolution module.
Further, the identity discriminator D is inputted in step 4idPerforming corresponding pixel subtraction on pixels in the image, and solving a square difference; then, a batch processing normalization layer (BN) and a full connection layer (FC) are carried out, and finally, a nonlinear function outputs classification probability; if the input and target are the same person, the tag is set to true, otherwise it is set to false.
Further, the attitude is input to the attitude discriminator D in step 5pdThe images and pose maps in (1) are cascaded along the channel dimension and then processed through 4 Conv-BN-ReLU sub-networks and nonlinear functions to obtain values in the interval [0, 1 ]]The confidence of (c).
Further, inputting the query graph and the graph to be queried to a decoding network in step 6, extracting features, calculating Euclidean distances of the features of the query graph and the graph to be queried, and sorting the Euclidean distances in an ascending order as the distances are more similar, and taking the results of Top1, Top5 and Top 10.
Has the advantages that: compared with the prior art, the invention has the following advantages: (1) when the pedestrian re-identification is finally carried out, only an encoder in the GAN network is used for extracting features, and extra attitude information and computing power are not needed;
(2) for the illumination change, the recognition rate of different postures of pedestrians with the same scene change and posture change is higher;
(3) different pedestrians have strong distinguishing force due to clothes similarity and the like;
(4) the generated graph is clearer and contains more detailed information.
Drawings
FIG. 1 is a diagram of the overall architecture of the process of the present invention;
FIG. 2 is a block diagram of an encoder network of the present invention;
FIG. 3 is a diagram of a decoding network model architecture of the present invention;
FIG. 4 is a graph of the results of the present invention.
Fig. 5 is a pedestrian search result diagram of the present invention.
Detailed Description
The present invention will be further illustrated with reference to the accompanying drawings and specific examples, which are carried out on the premise of the technical solution of the present invention, and it should be understood that these examples are only for illustrating the present invention and are not intended to limit the scope of the present invention.
The blocking indirect coupling GAN network consists of an encoder, a decoder and two discriminators;
as shown in fig. 1, the pedestrian re-identification method based on the blocking indirect coupling GAN network of the present invention has the following steps:
step 1: inputting a pedestrian image into an encoder in the GAN network;
step 2: training an encoder to obtain characteristics capable of distinguishing different pedestrians; in step 2, the encoder network structure is trained and is composed of a partitioned indirect coupling representation learning network;
and step 3: inputting the pedestrian picture again, and extracting the pedestrian features through an encoder; inputting a target attitude heat map, and obtaining attitude characteristics through an attitude characteristic extraction network; then combining the pedestrian feature vector, the attitude feature vector and the noise vector together in a channel dimension through concate operation, inputting the combined vectors into a decoder in a GAN network, and then training to generate a target attitude graph (namely a generated graph) of the pedestrian;
and 4, step 4: inputting the input pedestrian image, the target posture original image, and the generated image into an identity discriminator Did in the GAN network;
and 5: an attitude determination device D for inputting the input pedestrian image attitude map, target attitude original map, and generated map into GAN networkpdPerforming the following steps;
step 6: training the whole GAN network, finally utilizing the coding network in the GAN network to extract the characteristics of the query graph and the graph to be queried, and calculating the similarity of the characteristics for re-identification.
The specific implementation is as follows
The detailed steps of step 2 are as follows:
as shown in fig. 2, the pedestrian image is scaled to 384 × 128 size uniformly, and the image is subjected to image enhancement operations such as horizontal flipping and random erasing, and the processed image is input into the ResNet50 network in the encoder (the ResNet50 network is a part of the encoder, and the ResNet50 is the reference network framework of the encoder);
adopting ResNet50 as backbone network, and deleting original global average pool layer and network structure behind it; then, the resulting three-dimensional tensor is divided equally into six blocks in the horizontal direction before the pooling layer, and for each block, the average pooling and 1 × 1 convolution operations are performed; through the above operations, we can obtain a 256-dimensional column vector for each block; predicting a class label for each 256-dimensional vector of each block, and calculating cross entropy loss with a real class label;
in order to obtain more and richer features and ensure that the feature redundancy between different blocks is less and a complementary effect can be achieved, the cosine similarity of any two eigenvectors is calculated by adopting NCM loss, in order to ensure that any two eigenvectors are not similar, the blocking indirect coupling loss is provided, the calculated similarity is enabled to be L1 loss with zero vector measurement, in addition, in order to avoid the situation that the network weight matrix is a zero matrix when the calculated similarity is close to 0, the network weight matrix is subjected to L2 regularization, and thus the segment independent features are obtained.
Detailed description of step 3
The decoding network Decoder of the Decoder is an upsampling process. The method specifically comprises the following steps: the pedestrian features, the features of the target attitude and the noise vectors are fused to be used as the input of a decoder, and then the up-sampling operation is carried out to decode the input image size (namely, the image is generated).
The sub-module of the general decoder up-sampling is composed of Relu-Transconv-Norm network structure, namely, a deconvolution module; in order to fully utilize the characteristic information of a decoding layer, a deconvolution module and a Bilinear interpolation module are fused, the Bilinear interpolation module consists of a Biliner-Cov network, a 1x1 convolution layer is added to obtain the same number of characteristic channels as the deconvolution module, and only two layers of interpolation modules are added to realize the rapid convergence of the network.
As shown in fig. 3, according to the number of upsampling operations, we name the feature vector after the upsampling operation as a layer 1 feature vector, a layer 2 feature vector, a layer 3 feature vector, and a layer 4 feature vector in sequence to generate a map.
Specifically, the fused features are subjected to deconvolution module to obtain feature vectors of a 1 st layer, then the feature vectors of a 2 nd layer are obtained through the deconvolution module, the deconvolution module and the bilinear interpolation module are fused in the up-sampling operation of the feature vectors of the 2 nd layer to obtain feature vectors of a 3 rd layer, the feature vectors of a 4 th layer are obtained through the up-sampling operation of the feature vectors of the 3 rd layer through the fusion module, and finally the generated image is obtained through the feature vectors of the 4 th layer through the deconvolution module. And in the third and fourth times of upsampling, a fusion module is adopted instead of all layers, so that the quality of the generated image is effectively improved, and the convergence speed of the network is not reduced.
Detailed steps of step 4 and step 5: and judging the real image and generating an image.
The invention comprises two discriminators, e.g. FIG. 1(D) and (e), an identity discriminator DidAn attitude discriminator Dpd. Identity discriminator DidThe input image is encoded using ResNet50 as the backbone network, but the weight parameters are not shared with the Encoder network. Corresponding pixel subtraction is performed on the obtained pixels in the input image and the target image (or generation map), and then the squared difference is found. Then, a batch processing normalization layer (BN) and a full connection layer (FC) are carried out, and finally, a nonlinear function outputs classification probability. If the input and target are the same person, the tag is set to true, otherwise it is set to false. According to these operations, the distance between different gestures of the same ID can be reduced.
Attitude discriminator DpdFor determining whether the generated map has the same pose as the target. The gesture discriminator uses PatchGAN. The input image and pose graph are first cascaded along the channel dimension and then passed through 4 Conv-BThe N-ReLU sub-network and the non-linear function are processed to obtain values in the interval [0, 1 ]]Wherein the confidence map represents a degree of match between the input image and the pose map.
Detailed step of step 6:
training the whole GAN network to ensure that the network can obtain the characteristics of distinguishing different pedestrians and identifying the same pedestrian in different postures; and finally, inputting the query graph and the graph to be queried to a decoding network, extracting features, calculating Euclidean distances of the features of the query graph and the graph to be queried, sequencing the Euclidean distances in an ascending order when the distances are more similar, and taking the results of Top1, Top5 and Top 10.
Fig. 4 shows the result of the map generation of the present invention, the first line is the input image of the pedestrian, the second line is the target image to be generated by the network, the third line is the result of the map generation of the current better GAN network, and the fourth line is the result map of the present invention.
Fig. 5 shows Top10 results of a pedestrian search on a Market1501 data set, using a model trained by the present invention. The column to the left of the dotted line is the query graph and the right is the top10 searched images with the best image of the query graph.
The foregoing is only a preferred embodiment of the present invention, and it should be noted that, for those skilled in the art, various modifications and decorations can be made without departing from the principle of the present invention, and these modifications and decorations should also be regarded as the protection scope of the present invention.
Claims (8)
1. The pedestrian re-identification method based on the blocking indirect coupling GAN network is characterized by comprising the following steps: comprises the following steps of (a) carrying out,
step 1: inputting a pedestrian image into an encoder in the GAN network; the GAN network consists of an encoder, a decoder and two discriminators; two discriminators being identity discriminators DidAnd attitude discriminator Dpd;
Step 2: training an encoder to obtain characteristics capable of distinguishing different pedestrians;
and step 3: inputting the pedestrian picture again, and extracting the pedestrian features through an encoder; inputting a target attitude heat map, and obtaining attitude characteristics through an attitude characteristic extraction network; then combining the pedestrian feature vector, the attitude feature vector and the noise vector together in a channel dimension through concatee operation, inputting the combined pedestrian feature vector, attitude feature vector and noise vector into a decoder in a GAN network, and then training the combined pedestrian feature vector, attitude feature vector and noise vector to obtain a generated image;
and 4, step 4: an identity discriminator D for inputting the input pedestrian image, the target posture original image and the generated image into the GAN networkidPerforming the following steps;
and 5: an attitude discriminator D for inputting the input pedestrian image attitude map, the target attitude map and the generated map into the GAN networkpdPerforming the following steps;
step 6: training the whole GAN network, extracting the characteristics of the query graph and the graph to be queried by using the coding network in the GAN network, and calculating the similarity of the characteristics for re-recognition.
2. The pedestrian re-identification method based on the blocking indirect coupling GAN network as claimed in claim 1, wherein: in step 2, the Encoder is composed of a partitioned indirect coupling characterization learning network (namely, an encoding network Encoder), and step 2 comprises:
step 2-1: uniformly scaling the pedestrian images to 384 × 128, and performing image enhancement operations such as horizontal turning, random erasing and the like on the images;
step 2-2: inputting the graph obtained in the step 2-1 into a coding network for training, horizontally dividing the obtained feature vectors into 6 blocks, obtaining 256-dimensional column vectors for each block, and calculating cross entropy loss with a real class label;
step 2-3: and (4) obtaining the vector of each image block obtained in the step 2-1, calculating the cosine similarity of any two characteristic vectors, and measuring the L1 loss with the zero vector.
3. The pedestrian re-identification method based on the blocking indirect coupling GAN network as claimed in claim 2, wherein: the encoding network Encoder is composed of a reference network and two loss functions, and is specifically described as follows:
for step 2-2, for each of the 6 blocks, performing an average pooling and 1x1 convolution operation; obtaining a 256-dimensional column vector for each block;
and 2-3, performing L2 regularization on the network weight matrix when the cosine similarity is close to zero, thereby obtaining the characteristic of segment independence.
4. The pedestrian re-identification method based on the blocking indirect coupling GAN network as claimed in claim 1, wherein: the decoding network Decoder of the Decoder in step 3 is an up-sampling process, and step 3 includes:
step 3-1: inputting the fused features, and obtaining a feature vector of a layer 1 through the operation of a deconvolution module;
step 3-2: inputting the output of the 1 st layer into a deconvolution module to obtain a feature vector of the 2 nd layer;
step 3-3: the up-sampling operation of the feature vector of the layer 2 integrates a deconvolution module and a bilinear interpolation module, so that a layer 3 feature vector is obtained;
step 3-4: the feature vector of the layer 3 is subjected to the up-sampling operation of the fusion module to obtain a feature vector of the layer 4;
step 3-5: and the feature vector of the 4 th layer is subjected to a deconvolution module to obtain a generated graph.
5. The pedestrian re-identification method based on the blocking indirect coupling GAN network as claimed in claim 4, wherein: the acquisition process on the decoding network Decoder is as follows: the network input layer fuses the pedestrian characteristics, the target attitude characteristics and the combined characteristics of the noise vectors, then carries out up-sampling operation and decodes the combined characteristics into the size of an input graph;
the sub-modules sampled by the decoder are composed of a deconvolution module and a Bilinear interpolation module, the Bilinear interpolation module is composed of a Biliner-Cov network, and a 1x1 convolution layer is added to obtain the same number of characteristic channels as the deconvolution module.
6. The pedestrian re-identification method based on the blocking indirect coupling GAN network as claimed in claim 1, wherein: in step 4, the identity is input into the discriminatorDidPerforming corresponding pixel subtraction on pixels in the image, and solving a square difference; then, a batch processing normalization layer (BN) and a full connection layer (FC) are carried out, and finally, a nonlinear function outputs classification probability; if the input and target are the same person, the tag is set to true, otherwise it is set to false.
7. The pedestrian re-identification method based on the blocking indirect coupling GAN network as claimed in claim 1, wherein: in step 5, an input posture discriminator DpdThe images and pose maps in (1) are cascaded along the channel dimension and then processed through 4 Conv-BN-ReLU sub-networks and nonlinear functions to obtain values in the interval [0, 1 ]]The confidence of (c).
8. The pedestrian re-identification method based on the blocking indirect coupling GAN network as claimed in claim 1, wherein: inputting the query graph and the graph to be queried to a decoding network in step 6, extracting features, calculating Euclidean distances of the features of the query graph and the graph to be queried, sequencing the Euclidean distances in an ascending order when the distances are more similar, and taking the results of Top1, Top5 and Top 10.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202010218063.XA CN113449552A (en) | 2020-03-25 | 2020-03-25 | Pedestrian re-identification method based on blocking indirect coupling GAN network |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202010218063.XA CN113449552A (en) | 2020-03-25 | 2020-03-25 | Pedestrian re-identification method based on blocking indirect coupling GAN network |
Publications (1)
Publication Number | Publication Date |
---|---|
CN113449552A true CN113449552A (en) | 2021-09-28 |
Family
ID=77807554
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN202010218063.XA Pending CN113449552A (en) | 2020-03-25 | 2020-03-25 | Pedestrian re-identification method based on blocking indirect coupling GAN network |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN113449552A (en) |
Cited By (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN117612266A (en) * | 2024-01-24 | 2024-02-27 | 南京信息工程大学 | Cross-resolution pedestrian re-identification method based on multi-scale image and feature layer alignment |
Citations (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN108537136A (en) * | 2018-03-19 | 2018-09-14 | 复旦大学 | The pedestrian's recognition methods again generated based on posture normalized image |
CN109886251A (en) * | 2019-03-11 | 2019-06-14 | 南京邮电大学 | A kind of recognition methods again of pedestrian end to end guiding confrontation study based on posture |
CN110427813A (en) * | 2019-06-24 | 2019-11-08 | 中国矿业大学 | Pedestrian's recognition methods again based on the twin production confrontation network that posture instructs pedestrian image to generate |
WO2019227294A1 (en) * | 2018-05-28 | 2019-12-05 | 华为技术有限公司 | Image processing method, related device and computer storage medium |
-
2020
- 2020-03-25 CN CN202010218063.XA patent/CN113449552A/en active Pending
Patent Citations (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN108537136A (en) * | 2018-03-19 | 2018-09-14 | 复旦大学 | The pedestrian's recognition methods again generated based on posture normalized image |
WO2019227294A1 (en) * | 2018-05-28 | 2019-12-05 | 华为技术有限公司 | Image processing method, related device and computer storage medium |
CN109886251A (en) * | 2019-03-11 | 2019-06-14 | 南京邮电大学 | A kind of recognition methods again of pedestrian end to end guiding confrontation study based on posture |
CN110427813A (en) * | 2019-06-24 | 2019-11-08 | 中国矿业大学 | Pedestrian's recognition methods again based on the twin production confrontation network that posture instructs pedestrian image to generate |
Non-Patent Citations (2)
Title |
---|
YUE ZHANG; YI JIN; JIANQIANG CHEN; SHICHAO KAN; YIGANG CEN; QI CAO: "PGAN: Part-Based Nondirect Coupling Embedded GAN for Person Reidentification", 《IEEE MULTIMEDIA》 * |
冯敏;张智成;吕进;余磊;韩斌;: "基于生成对抗网络的跨模态行人重识别研究", 现代信息科技, no. 04 * |
Cited By (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN117612266A (en) * | 2024-01-24 | 2024-02-27 | 南京信息工程大学 | Cross-resolution pedestrian re-identification method based on multi-scale image and feature layer alignment |
CN117612266B (en) * | 2024-01-24 | 2024-04-19 | 南京信息工程大学 | Cross-resolution pedestrian re-identification method based on multi-scale image and feature layer alignment |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN107844779B (en) | Video key frame extraction method | |
Qu et al. | RGBD salient object detection via deep fusion | |
Zhang et al. | Joint human detection and head pose estimation via multistream networks for RGB-D videos | |
CN107239730B (en) | Quaternion deep neural network model method for intelligent automobile traffic sign recognition | |
CN110717411A (en) | Pedestrian re-identification method based on deep layer feature fusion | |
CN114783003B (en) | Pedestrian re-identification method and device based on local feature attention | |
Fooladgar et al. | Multi-modal attention-based fusion model for semantic segmentation of RGB-depth images | |
CN114187665B (en) | Multi-person gait recognition method based on human skeleton heat map | |
Liu et al. | Pedestrian detection algorithm based on improved SSD | |
CN110334622B (en) | Pedestrian retrieval method based on adaptive feature pyramid | |
Saqib et al. | Person head detection in multiple scales using deep convolutional neural networks | |
Wang et al. | MCF3D: Multi-stage complementary fusion for multi-sensor 3D object detection | |
CN111325169A (en) | Deep video fingerprint algorithm based on capsule network | |
Cho et al. | Semantic segmentation with low light images by modified CycleGAN-based image enhancement | |
Tian et al. | Snowflake removal for videos via global and local low-rank decomposition | |
Tao et al. | CENet: A channel-enhanced spatiotemporal network with sufficient supervision information for recognizing industrial smoke emissions | |
Lejbolle et al. | Attention in multimodal neural networks for person re-identification | |
CN113221770A (en) | Cross-domain pedestrian re-identification method and system based on multi-feature hybrid learning | |
Nguyen et al. | A robust triangular sigmoid pattern-based obstacle detection algorithm in resource-limited devices | |
Rahman et al. | Pedestrian Detection in Thermal Images Using Deep Saliency Map and Instance Segmentation | |
Xie et al. | Learning visual-spatial saliency for multiple-shot person re-identification | |
Nie et al. | Multiple person tracking by spatiotemporal tracklet association | |
CN113449552A (en) | Pedestrian re-identification method based on blocking indirect coupling GAN network | |
Huang et al. | Multi‐class obstacle detection and classification using stereovision and improved active contour models | |
Ma et al. | MSFNET: multi-stage fusion network for semantic segmentation of fine-resolution remote sensing data |
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 |