Research On Person Re-identification Methods Based On Dual-stream Feature Fusion And Two-level Salient Feature Complementary

Person re-identification(Re-ID)is a computer vision technique that retrieves pedestrians with the same identity in a specific video and image set by giving a target pedestrian image.Traditional person Re-ID methods are mainly based on information such as texture features and color of pedestrian images to accomplish matching of pedestrian images.In recent years,with the development of deep learning technology,the person Re-ID methods based on deep learning has greatly improved the accuracy of person Re-ID.However,in practical scenarios,person Re-ID tasks still face various challenges such as pose variations,pedestrian misalignment,occlusions,cluttered backgrounds,and so on.In addition,similar color of different pedestrian’s dress may also seriously affect the performance of person Re-ID.Therefore,it is important to extract robust and discriminative features in pedestrian images for person Re-ID tasks.To this end,the main work of this paper is as follows:(1)In order to further weaken the effect of similarity in dress color among different pedestrians on the accuracy of person Re-ID,we propose a dual-stream feature fusion network(DSFF-Net).Specifically,we first design a dual-stream network to extract RGB global features,grayscale global features and local features of pedestrian images to increase the richness of pedestrian representations.Second,we design a channel attention module to guide the network to focus on the salient features of pedestrians.Then,we design an embedding mixed pooling operation to obtain more discriminative global features of pedestrians by combining the outputs of the global average pooling(GAP)and the global max pooling(GMP),which can also remove redundant information and increase the robustness of pedestrian representations.Finally,we design a fine-grained local feature embedding fusion operation to obtain more discriminative local features of pedestrians by embedding and fusing fine-grained local features of RGB pedestrian images and GRAY pedestrian images to increase the richness of pedestrian representations.(2)In order to make the network focus on the most salient discriminative features and secondary salient discriminative features simultaneously,we propose a two-level salient feature complementary network(TSFC-Net)to extract the most salient discriminative features and the secondary salient discriminative features of pedestrian images for person Re-ID.Specifically,we first design a spatial attention module and a channel attention module,which are embedded into the backbone network to extract the most salient discriminative features of pedestrians.Then,we design a secondary salient feature mining module to extract the secondary salient discriminative features of pedestrians.Since the final features of pedestrian images fuse the most salient discriminative features and the secondary salient discriminative features,TSFC-Net can significantly improve the richness and discrimination capability of pedestrian representations.(3)We conduct extensive experiments on the Market-1501,Duke MTMC-re ID,and CUHK03 datasets,and the experimental results indicate that DSFF-Net and TSFC-Net have better performance compared with most of the state-of-the-art person Re-ID methods.