Cloth-Changing Person Re-Identification Based On Deep Learning

In recent years,due to the emergence of large-scale open-source datasets and the rapid development of deep learning networks,person re-identification technology has achieved many excellent research results.However,many existing methods rely to a large extent on the clothing appearance of pedestrians in a short period of time to conduct retrieval,and do not consider the high probability of a pedestrian changing clothes over a long period of time.Clothing changes have led to a performance decline in existing pedestrian re-identification models.However,in real life,a pedestrian has a high probability of changing clothes over a relatively long period of time.Therefore,researching person re-identification in clothing change scenarios is necessary.For the cloth-changing person re-identification problem,this thesis has done the following work:(1)A cloth-changing person re-identification method based on fine-grained pairwise interaction network is proposed to reduce the interference of clothing information on the network model and further extract more robust pedestrian features.First,by introducing spatial attention modules to highlight the importance of specific channels,valuable fine-grained features are selected to reduce the impact of clothing factors on the model’s performance.Then,a paired interaction network is designed to capture similarity correlation regions in pedestrian image pairs with the same identity but different clothing and learn consistent features before and after clothing changes.On this basis,more prominent pedestrian features are obtained through attention pooling,and the mean square error and triplet loss functions are used to constrain this feature to improve the retrieval ability of the model.(2)A cloth-changing person re-identification method based on generative adversarial networks is proposed.Generating more training data while separating the identity-related and unrelated features of pedestrians to make the identity-related features more robust.First,two encoders are used to extract identity-related and irrelevant features of pedestrians,respectively.Then,the generator is used for image self-reconstruction,cross-reconstruction of pedestrian images with the same identity,and cross-identity reconstruction of pedestrian images.Different loss functions are introduced to improve the image generation quality.The identity discriminator is used to classify images and regulate the feature learning process.Finally,the model is trained by minimizing the weighted loss function to improve the re-identification ability of the model while constraining the image generation quality of the generator,making the model more robust.(3)Experiments were conducted on three public open source datasets,PRCC,Celeb-re ID,and Celeb-re ID-light,to evaluate the proposed methods.The experimental results show that the cloth-changing person re-identification method based on finegrained paired interaction networks achieved Rank-1 accuracies of 69.5%,65.2%,and48.0% and mean average precisions of 61.8%,19.3%,and 26.3% on the three datasets,respectively.The cloth-changing person re-identification method based on generative adversarial networks achieved Rank-1 accuracies of 51.2%,64.1%,and 35.8% and mean average precisions of 54.8%,16.4%,and 18.6% on the three datasets,respectively.The experimental results demonstrate that compared to other state-of-theart methods,the methods proposed in this thesis have certain advantages when pedestrian clothing changes.