Research On Person Re-identification Based On Multi-scale Features And Deep Multi-branch Network

With the rapid breakthrough of deep learning technology and people's urgent demand for intelligent security technology,the research work of person re-idtification has achieved remarkable development.However,in the actual shooting scene,because of the change of posture,angle,lighting conditions,and the emergence of the redundant of pedestrian background information and occlusion problem,further breakthroughs in person re-identification technology are restricted,and further research is still needed.At present,the construction of new deep learning models to extract high robustness and high discrimination pedestrian feature representation is still a key research direction to solve various person re-identification problems.In this paper,through multi-scale feature learning,multi-granularity learning and other methods,new network model is proposed and the feature extraction branch is adjusted to obtain more comprehensive and discriminant features,so as to improve the accuracy and robustness of person re-identification.The important work and innovation achievements of this paper include:(1)A multi-scale attention network based on multi-feature fusion is proposed for person re-identification task.The convolutional attention module is introduced in this method,which makes the network focus more on the key features conducive to reidentification and increases the learning ability of the network.Secondly,the channel attention module is organically integrated with the pooling fusion strategy(GAP+GMP)to obtain the feature description with high discrimination and robustness,thus improving the accuracy of person re-identification.(2)A two-stream person re-identification network based on multi-scale feature learning is proposed.Different from the work mentioned above,this method considers how to extract a wider range of multi-scale features to enhance feature representation.In this method,the learning structure of multi-scale feature learning is built on the structure of the two-stream network,which greatly expands the sampling range of features,and realizes the effective fusion of global features and local features with different granularity,so that the network can obtain more discernible features and improve the accuracy of person re-identification.The first algorithm tends to use attention mechanism to focus on key pedestrian features at different scales,while the second algorithm focuses on using two-stream multi-branch structure to expand the sampling range of multi-scale features.