Research On Person-Re-Identification Based On Spatial And Channel Attention Mechanism

Person re-identification is an important research task in the field of computer vision,which aims to identify the same person in the images captured by cameras with non-overlapping perspectives.It has broad application prospects in the field of security monitoring.In recent years,person re-identification has been greatly promoted because of the deep learning technology and has attracted widespread attention.To make full use of the information in pedestrian images,this paper proposes a method for person re-identification based on spatial and channel attention mechanisms.The spatial attention model is based on the idea of non-local mean filtering operation.When calculating the output of each pixel position,it no longer calculates only with the neighborhood,but calculates the correlation with all positions in the image.It regards the correlation as a weight to indicate the similarity between other locations and the current location.By paying attention to all locations in the feature map and using their weights in the embedding space,the model can capture the spatial dependences between locations.In the channel attention model,the global average pooling and global maximum pooling operations are employed to aggregate the spatial information of the feature map,which removes the correlation of the spatial dimensions.The model captures the dependencies between the channels through the bottleneck structure and add a weight to each channel to represent the importance of the channel.A feature vector extraction module is designed at the end of the network to combine two different features to obtain a feature vector with better generalization ability.In addition,this paper also combines global features and local features,which learns the overall attributes of the entire pedestrian image and the most discriminative features of the local area respectively.In this paper,extensive experiments are performed on,three public datasets:Market1501,DukeMTMC-reID and CUHK03.The experimental results show that the proposed method achieves better performance than the state-of-the-art methods.The effectiveness of each module in the network has also been proved in comparative experiments.