Method and apparatus for generating target re-recognition model and re-recognizing target

Patent metadata

Field	Value
Publication number	US-11915484-B2
Application number	US-202117304296-A
Country	US
Kind code	B2
Filing date	Jun 17, 2021
Priority date	Oct 23, 2020
Publication date	Feb 27, 2024
Grant date	Feb 27, 2024

Abstract

A method, an apparatus, device and a storage medium for generating a target re-recognition model are provided. The method may include: acquiring a set of labeled samples, a set of unlabeled samples and an initialization model obtained through supervised training; performing feature extraction on each sample in the set of the unlabeled samples by using the initialization model; clustering features extracted from the set of the unlabeled samples by using a clustering algorithm; assigning, for each sample in the set of the unlabeled samples, a pseudo label to the sample according to a cluster corresponding to the sample in a feature space; and mixing a set of samples with a pseudo label and the set of the labeled samples as a set of training samples, and performing supervised training on the initialization model to obtain a target re-recognition model.

First claim

What is claimed is:

1. A method for generating a target re-recognition model, the method comprising:
acquiring a set of labeled samples, a set of unlabeled samples and an initialization model obtained through supervised training;
performing feature extraction on each sample in the set of the unlabeled samples by using the initialization model;
clustering features extracted from the set of the unlabeled samples by using a clustering algorithm;
assigning, for each sample in the set of the unlabeled samples, a pseudo label to the each sample according to a cluster corresponding to the each sample in a feature space; and
mixing a set of samples with a pseudo label and the set of the labeled samples as a set of training samples, and performing the supervised training on the initialization model to obtain a target re-recognition model, wherein mixing the set of samples with the pseudo label and the set of the labeled samples as the set of training samples, and performing the supervised training on the initialization model to obtain the target re-recognition model, comprises:
performing feature extraction on each sample in the set of the labeled samples by using the initialization model;
fusing, by using a graph convolutional neural network, a feature of each sample in the set of the samples with the pseudo label and a feature of each sample in the set of the labeled samples to obtain a fused feature of each sample; and
training the initialization model based on the fused feature of each sample in the set of the samples with the pseudo label and the fused feature of each sample in the set of the labeled samples to obtain the target re-recognition model.

2. The method according to claim 1 , wherein the graph convolutional neural network comprises a first graph convolutional layer and a second graph convolutional layer, and wherein:
the first graph convolutional layer comprises at least one sample node representing a sample and at least one proxy node representing a set of samples, and sample nodes belonging to a given set of samples are unidirectionally connected to a given proxy node, and proxy nodes are interconnected, and each proxy node performs a weighted sum on sample features of sample nodes connected to the each proxy node to obtain a proxy feature of the each proxy node, and proxy features of all proxy nodes are fused through the first graph convolutional layer to obtain output features of the proxy nodes of the first graph convolutional layer; and
the second graph convolutional layer comprises at least one sample node representing a sample and at least one proxy node representing a set of samples, and sample nodes belonging to a given set of samples are bidirectionally connected to a given proxy node, and proxy nodes are interconnected, and the output features of the proxy nodes of the first graph convolutional layer are fused through the second graph convolutional layer to obtain an output feature of each sample node.

3. The method according to claim 2 , wherein the graph convolutional neural network comprises the first graph convolutional layer and at least one second graph convolutional layer, wherein an output of the first graph convolutional layer is used as an input of the second graph convolutional layer, and an output feature of each second graph convolutional layer is used as an input of a next second graph convolutional layer or an input of a classification layer of the initialization model.

4. The method according to claim 1 , wherein the method uses at least one set of labeled samples, and each of the at least one set of labeled samples is from one data source.

5. The method according to claim 1 , further comprising:
acquiring a to-be-recognized monitoring image; and
inputting the monitoring image into the target re-recognition model to generate a target re-recognition result.

6. An electronic device, comprising:
at least one processor; and
a memory storing instructions executable by the at least one processor, the instructions, when executed by the at least one processor, cause the at least one processor to perform operations comprising:
acquiring a set of labeled samples, a set of unlabeled samples and an initialization model obtained through supervised training;
performing feature extraction on each sample in the set of the unlabeled samples by using the initialization model;
clustering features extracted from the set of the unlabeled samples by using a clustering algorithm;
assigning, for each sample in the set of the unlabeled samples, a pseudo label to the each sample according to a cluster corresponding to the each sample in a feature space; and
mixing a set of samples with a pseudo label and the set of the labeled samples as a set of training samples, and performing the supervised training on the initialization model to obtain a target re-recognition model, wherein mixing the set of samples with the pseudo label and the set of the labeled samples as the set of training samples, and performing the supervised training on the initialization model to obtain the target re-recognition model, comprises:
performing feature extraction on each sample in the set of the labeled samples by using the initialization model;
fusing, by using a graph convolutional neural network, a feature of each sample in the set of the samples with the pseudo label and a feature of each sample in the set of the labeled samples to obtain a fused feature of each sample; and
training the initialization model based on the fused feature of each sample in the set of the samples with the pseudo label and the fused feature of each sample in the set of the labeled samples to obtain the target re-recognition model.

7. The electronic device according to claim 6 , wherein the graph convolutional neural network comprises a first graph convolutional layer and a second graph convolutional layer, and wherein:
the first graph convolutional layer comprises at least one sample node representing a sample and at least one proxy node representing a set of samples, and sample nodes belonging to a given set of samples are unidirectionally connected to a given proxy node, and proxy nodes are interconnected, and each proxy node performs a weighted sum on sample features of sample nodes connected to the each proxy node to obtain a proxy feature of the each proxy node, and proxy features of all proxy nodes are fused through the first graph convolutional layer to obtain output features of the proxy nodes of the first graph convolutional layer; and
the second graph convolutional layer comprises at least one sample node representing a sample and at least one proxy node representing a set of samples, and sample nodes belonging to a given set of samples are bidirectionally connected to a given proxy node, and proxy nodes are interconnected, and the output features of the proxy nodes of the first graph convolutional layer are fused through the second graph convolutional layer to obtain an output feature of each sample node.

8. The electronic device according to claim 7 , wherein the graph convolutional neural network comprises the first graph convolutional layer and at least one second graph convolutional layer, wherein an output of the first graph convolutional layer is used as an input of the second graph convolutional layer, and an output feature of each second graph convolutional layer is used as an input of a next second graph convolutional layer or an input of a classification layer of the initialization model.

9. The electronic device according to claim 6 , wherein the operations use at least one set of labeled samples, and each of the at least one set of labeled samples is from one data source.

10. The electronic device according to

Assignees

• Beijing Baidu Netcom Sci & Tech Co Ltd

Inventors

• Wang Zhigang
• Wang Jian
• Ding Errui
• Sun Hao

Classifications

• G06N3/0895

  Weakly supervised learning, e.g. semi-supervised or self-supervised learning · CPC title

• G06N3/09

  Supervised learning · CPC title

• G06N3/0464

  Convolutional networks [CNN, ConvNet] · CPC title

• G06V20/52 Primary

  Surveillance or monitoring of activities, e.g. for recognising suspicious objects (recognising microscopic objects G06V20/69) · CPC title

• G06F18/214

  Generating training patterns; Bootstrap methods, e.g. bagging or boosting · CPC title