Cross-domain pedestrian re-identification method and system based on deep mutual learning

Abstract

The invention relates to a cross-domain pedestrian re-identification method and system based on deep mutual learning. The method comprises the following steps: pre-training an initial network in a source domain to obtain a basic network; constructing a three-branch mutual learning network framework and training in a target domain, wherein each branch comprises a main network PNet, an average network Mean Net and a difficult sample screening classifier FNet; in each branch, the FNet inputs samples into PNet training from easy to difficult, and meanwhile, the FNet adopts triple loss training; the PNet obtains a simple sample judged by the FNet, transmits the simple sample to the Mean Net, and judges a soft pseudo tag corresponding to the sample; when the cosine similarity between the sample features, extracted by the Mean Net, of any two branches is smaller than a set threshold value, the sample is determined as a high-confidence-coefficient sample, and the high-confidence-coefficient sample is used for updating the network in the third branch; and after iteration training is finished, Mean Net with the best performance in the three branches is selected for cross-domain pedestrian re-identification. The method and the system are beneficial to improving the accuracy of cross-domain pedestrian re-identification.

Description

technical field [0001] The invention belongs to the technical field of computer vision, and in particular relates to a cross-domain pedestrian re-identification method and system based on deep mutual learning. Background technique [0002] Person re-identification technology is a research hotspot in computer vision, artificial intelligence and other disciplines. Pedestrian re-identification is mainly used to determine whether pedestrian pictures taken at different times and under different cameras are the same pedestrian. With the development of digital cities and smart cities, pedestrian re-identification has been widely used in video surveillance, criminal investigation security, intelligent security and other fields. [0003] The purpose of pedestrian re-identification is to accurately find specific pedestrians from massive video and picture data, so that computers can quickly and accurately retrieve target pedestrians in complicated surveillance videos and pictures inst...

AI Technical Summary

Problems solved by technology

However, the traditional collaborative learning method consists of two models updating each other, which will show convergence after multiple rounds of iterations, thus reducing the model's ability to judge noise labels

[0007] In addition, in the cross-domain person re-identification method, the model trained in one dataset is directly used in another dataset, which often suffers from a large-scale drop in accuracy.

One of the reasons is that the generalization of unsupervised learning models is poor, so the models cannot be robust to different datasets

Another reason is that the datasets for pedestrian re-identification are different due to different shooting times, locations, angles, etc., and the feature distribution of samples between different datasets is different, resulting in domain gaps in different datasets, which affect the pedestrian re-identification model. performance

Images