Unsupervised cross-domain pedestrian re-identification method based on progressive enhanced self-learning

Abstract

The invention discloses an unsupervised cross-domain pedestrian re-identification method based on progressive enhanced self-learning. The unsupervised cross-domain pedestrian re-identification methodcomprises the following steps: firstly, extracting initial features of a label-free target domain data set from a network model obtained by training on a labeled original domain data set; calculatingsimilarity score matrices, utilizing HDBSCAN clustering to give a pseudo class label to the target domain data, and performing model retraining by using a loss function based on Triplet; and then carrying out retraining, two-stage alternate loop learning and mutual correction again on the network model with the clustering center feature initialization classification layer by using a Softmax loss function, and finally extracting discrimination features of the network model as feature representation of the picture to carry out feature-level comparison. By adopting the unsupervised cross-domain pedestrian re-identification method provided by the invention, the convergence capability of the network model and the information representation capability of the model can be gradually enhanced, andthe generalization capability of the network model on the label-free target domain data can be further improved, so that the unsupervised cross-domain pedestrian re-identification precision can be improved.

Description

technical field [0001] The invention relates to a pedestrian re-identification method, in particular to an unsupervised cross-domain pedestrian re-identification method with progressive enhancement self-learning, and belongs to the technical field of image processing. Background technique [0002] Pedestrian re-identification refers to: given a camera to be retrieved target pedestrians, locate them under different cameras, that is, to confirm whether the target pedestrians appear under other cameras one by one. [0003] Pedestrian re-identification has very important practical significance in the fields of video surveillance, security protection, and auxiliary investigation. [0004] In recent years, with the rapid development of deep learning, many supervised pedestrian re-identification works have achieved rapid progress, that is, deep neural network training is carried out on a large number of pedestrian datasets with identity information, and in the same scene. Pedestri...

AI Technical Summary

Problems solved by technology

However, labeling a large amount of data requires huge manpower and time costs, so many unsupervised learning methods have emerged, which can make full use of the extremely easy-to-obtain unlabeled data

Compared with labeled data, unlabeled data does not provide identity information, which makes network training lack guidance. Therefore, the effect of unsupervised learning method for pedestrian re-identification is very poor and cannot be applied in real life.

[0005] In order to solve the problem of poor pedestrian re-identification in unsupervised learning methods, many unsupervised cross-domain methods have proposed using both labeled data and unlabeled data.

However, directly applying the model trained on labeled data (original domain) to unlabeled data (target domain) will cause a huge performance deviation. The main problems are:

Images