Cross-Domain Person Re-Identification Methods Based On Deep Learning

Person re-identification mainly studies the matching problem of the same person image in different scenes and perspectives,and has broad application prospects in public security and other aspects.This thesis focuses on cross-domain person re-identification using deep learning algorithms.The goal of this research is to apply deep neural network models learned on source domain datasets to target domain datasets with distribution shifts,so as to alleviate the problem caused by the lack of people labels in the target domain.One of the main technical routes of the cross-domain person re-identification method based on deep learning is the self-training method,which focuses on dealing with the relationship between deep neural network model optimization and pseudo-label generation.Regarding the unreliable problem of pseudo-labels in self-training methods for cross-domain person re-identification,this thesis mainly studies and works on the following contents:(1)Cross-domain person re-identification method for low-quality blurred images.This thesis discusses the impact of person image quality on the generation of pseudolabels for cross-domain person re-identification from the perspective of data input.Through analysis,it is concluded that there are two key ideas for dealing with low-quality images: one is to construct a network model with more powerful and effective feature extraction,and the other is to estimate and process low-quality images.Based on these two ideas,this thesis constructs a more powerful cross-domain person re-identification encoding network using existing general person re-identification methods and proposes an adaptive person quality-weighted loss for unsupervised image quality-aware models based on autoencoder networks.Finally,the effectiveness of the proposed method is validated on public and self-made datasets.(2)A cross-domain person re-identification method based on deep clustering of the target domain.This thesis analyzes the inconsistency between the pseudo-label results of the target domain clustering algorithm and the real person labels,and studies the construction of a contrastive loss method that is resistant to pseudo-label noise in unreliable clustering results.To address the identity mis-separation phenomenon in clustering results,this thesis uses a multi-channel linear classifier initialized by class proxy inheritance to alleviate the problem of pulling closer person embedding vectors of different identities in the loss function.To address the identity mis-merging phenomenon in clustering results,this thesis improves the loss function using an implicit sample expansion method to alleviate the problem of pulling apart person embedding vectors of the same identity in the loss function.Finally,this thesis combines the two methods to comprehensively address the inconsistency between clustering results and real labels and conducts relevant experimental verification.(3)A cross-domain person re-identification method based on difficult sample sampling in the source domain.To improve the consistency between clustering results and real person labels during the training process of cross-domain person re-identification models,this thesis improves the hard sample mining method of triplet loss during the training process of source domain data.This thesis designs a hard sample mining strategy based on the difference between source domain data clustering results and real labels to reduce the distribution difference between source domain clustering results and real labels,thereby assisting cross-domain person re-identification models to fit key common invariant features that are both person-discriminative and clustering-friendly.