Pedestrian re-identification method and storage medium based on hybrid cluster center label learning

Abstract

The invention discloses a pedestrian re-identification method and a storage medium based on hybrid cluster center label learning. Firstly, network model parameters are initialized using labeled data, and initial cluster center labels are calculated, and then feature information of unlabeled data is extracted by using the network model; Then calculate the distance between the feature information of the unlabeled data and the cluster center and filter out the pseudo-label data of the preset ratio, and the remaining data is fuzzy label data, and generate the cluster center label as a guide label, and update the cluster center label in the memory ; Add pseudo-label data and fuzzy label data to the training samples in a small number of times, and retrain the deep neural network model. The present invention divides unlabeled data into pseudo-labeled data and fuzzy-labeled data by using a clustering method, and calculates the cluster center, and then uses various types of cluster centers to perform model classification optimization, fully utilizes various information, and effectively Improving the accuracy of person re-identification methods.

Description

technical field

[0001] The invention belongs to the technical field of pedestrian re-identification, and in particular relates to a pedestrian re-identification method and a storage medium based on hybrid cluster center label learning. Background technique

[0002] Person re-identification (Re-ID) is an image retrieval technique that retrieves whether a specific pedestrian exists in an image or video. Person re-identification based on domain adaptation is trained by labeled source domain data and unlabeled target domain data to detect specific pedestrians on the target domain. Person re-identification based on unsupervised domain adaptation is similar to the method of person re-identification based on domain adaptation, the difference is that the source domain data has no labels.

[0003] Existing UDA-based Re-ID training methods usually have the following two stages: (1) Supervised pre-training using labeled source domain data; (2) On the basis of pre-training, using unlab...

Examples

Embodiment 1

[0047] A person re-identification method based on hybrid cluster center label learning, such as figure 2 , image 3 shown, including the following steps:

[0048] Step S100: collecting a pedestrian re-identification training data set, and dividing the training data set into labeled data samples and unlabeled data samples;

[0049] Step S200: first use the labeled data to initialize the parameters of the deep neural network model, and calculate the initial cluster center label and store it in the memory, and then use the deep neural network model to extract the feature information of the unlabeled data;

[0050] Step S300: Use multiple cluster center label generation modules to calculate the distance between the feature information of the unlabeled data and the cluster center and filter out the pseudo-label data with a preset ratio, and the remaining data are fuzzy label data and pseudo-label data Generating the cluster center label with the fuzzy label data as a guide label, ...

Embodiment 2

[0059] This embodiment is optimized on the basis of Embodiment 1. In the step S400, unlabeled data is added in a small number of iterations, and a small number of unlabeled data is added to the deep neural network model for training. , each iteration adds less unlabeled data than the previous iteration.

[0060] Further, the added unlabeled data increases at a rate of 10% for the first 10 epochs, and increases at a rate of 1% in subsequent iterations.

[0061] The above-mentioned small number of iterations is introduced into the model training process because the growth rate of the number of unlabeled data in the training data will affect the recognition accuracy of the overall network model. Added to the network model for training, the unlabeled data added in each iteration will be less than the data added last time, and the overall capacity of the training samples will be increased, thereby enhancing the performance of the model. The first 10 epochs of the added unlabeled da...

Embodiment 3

[0064] This embodiment is optimized on the basis of embodiment 1 or 2, as figure 1 As shown, the front end of the deep neural network model is provided with a block layer, which is used to divide the pedestrian image into three equal parts; the deep neural network model includes a convolution layer, a batch normalization layer, an activation Function layer, residual module, neuron deactivation layer, fully connected layer.

[0065] Further, the deep neural network model adopts a 64-layer residual deep neural network model.

[0066] The front end of the deep convolutional neural network builds a block layer to divide the pedestrian image into three equal parts, which is beneficial to the subsequent extraction of the feature information of each component. The backbone part adopts a residual network model with a 64-layer structure. The structure is to stack the convolutional layer, batch normalization layer, activation function layer, residual module, neuron deactivation layer,...

technical field

[0001] The invention belongs to the technical field of pedestrian re-identification, and in particular relates to a pedestrian re-identification method and a storage medium based on hybrid cluster center label learning. Background technique

[0002] Person re-identification (Re-ID) is an image retrieval technique that retrieves whether a specific pedestrian exists in an image or video. Person re-identification based on domain adaptation is trained by labeled source domain data and unlabeled target domain data to detect specific pedestrians on the target domain. Person re-identification based on unsupervised domain adaptation is similar to the method of person re-identification based on domain adaptation, the difference is that the source domain data has no labels.

[0003] Existing UDA-based Re-ID training methods usually have the following two stages: (1) Supervised pre-training using labeled source domain data; (2) On the basis of pre-training, using unlab...

Examples

Embodiment 1

[0047] A person re-identification method based on hybrid cluster center label learning, such as figure 2 , image 3 shown, including the following steps:

[0048] Step S100: collecting a pedestrian re-identification training data set, and dividing the training data set into labeled data samples and unlabeled data samples;

[0049] Step S200: first use the labeled data to initialize the parameters of the deep neural network model, and calculate the initial cluster center label and store it in the memory, and then use the deep neural network model to extract the feature information of the unlabeled data;

[0050] Step S300: Use multiple cluster center label generation modules to calculate the distance between the feature information of the unlabeled data and the cluster center and filter out the pseudo-label data with a preset ratio, and the remaining data are fuzzy label data and pseudo-label data Generating the cluster center label with the fuzzy label data as a guide label, ...

Embodiment 2

[0059] This embodiment is optimized on the basis of Embodiment 1. In the step S400, unlabeled data is added in a small number of iterations, and a small number of unlabeled data is added to the deep neural network model for training. , each iteration adds less unlabeled data than the previous iteration.

[0060] Further, the added unlabeled data increases at a rate of 10% for the first 10 epochs, and increases at a rate of 1% in subsequent iterations.

[0061] The above-mentioned small number of iterations is introduced into the model training process because the growth rate of the number of unlabeled data in the training data will affect the recognition accuracy of the overall network model. Added to the network model for training, the unlabeled data added in each iteration will be less than the data added last time, and the overall capacity of the training samples will be increased, thereby enhancing the performance of the model. The first 10 epochs of the added unlabeled da...

Embodiment 3

[0064] This embodiment is optimized on the basis of embodiment 1 or 2, as figure 1 As shown, the front end of the deep neural network model is provided with a block layer, which is used to divide the pedestrian image into three equal parts; the deep neural network model includes a convolution layer, a batch normalization layer, an activation Function layer, residual module, neuron deactivation layer, fully connected layer.

[0065] Further, the deep neural network model adopts a 64-layer residual deep neural network model.

[0066] The front end of the deep convolutional neural network builds a block layer to divide the pedestrian image into three equal parts, which is beneficial to the subsequent extraction of the feature information of each component. The backbone part adopts a residual network model with a 64-layer structure. The structure is to stack the convolutional layer, batch normalization layer, activation function layer, residual module, neuron deactivation layer,...

Images