Pseudo label distribution method and system for pedestrian re-identification generated data

Abstract

The invention provides a pseudo label distribution method and system for pedestrian re-identification generated data. The method comprises the following steps: generating label-free data on the basis of a pedestrian re-identification real data set by using a deep convolution generative adversarial network method, and forming a label-free data set; distributing virtual labels for label-free data in the label-free data set according to the dynamic label distribution, and randomly setting the weight of a certain number of virtual labels to be 0 by setting a sparsification factor to obtain sparse regularized multi-pseudo labels; predicting by utilizing Softmax to obtain the probability that the pedestrian belongs to a certain predefined class; obtaining an optimized cross entropy loss function according to the label distribution and the probability; combining the label-free data set with the sparse regularization multi-pseudo labels, fusing the label-free data set with the real data set to form a training sample, and training the training sample according to a loss function to obtain a pedestrian re-identification model. Therefore, the over-fitting phenomenon of the model on the characteristics of some predefined classes is avoided, and the generalization ability of the model is improved.

Description

technical field [0001] The invention relates to the technical field of image recognition, in particular to a pseudo-label distribution method and system for generated data of pedestrian re-identification. Background technique [0002] Pedestrian re-identification is an important research topic in intelligent surveillance systems, and its purpose is mainly to identify the same pedestrian under non-overlapping camera views across regions. Pedestrian re-identification is a challenging topic, and its performance is often affected by factors such as pedestrian posture, illumination changes, pedestrian occlusion or misalignment. At present, person re-identification often uses convolutional neural network combined with supervised learning to obtain a discriminative model. The method of supervised learning requires a large number of training data sets. However, the current pedestrian re-identification data set is far from enough compared to the large data set of ImageNet. If the d...

AI Technical Summary

Problems solved by technology

However, this method will cause the following problems: when the clothing color in the original data is mostly dark, the generated data will tend to generate pedestrians with dark clothes

Due to this method, when MpRL dynamically assigns predetermined class weights, it will tend to assign greater weights to these predefined classes with dark clothes, so that light-colored clothes features will be ignored in the training process, and the generalization of the model will be greatly reduced

Images