Research On Global Feature Enhancement Algorithm In Person Re-identification

With the development of the security industry,surveillance cameras are widely used in our daily lives.Increasing video data and changing application scenarios have brought new requirements to the field of computer vision.Person re-identification is an important technology in the field of security surveillance,and it aims to match pedestrians from the non-overlapping surveillance cameras.Feature representation is a challenging part of person re-identification.It improves the accuracy of person re-identification by extracting strong and robust image features.This dissertation mainly researches on the feature extraction algorithm based on deep neural networks for person re-identification,and proposes a global feature enhancement model.The model enhances the global feature from both the spatial and channel dimensions of the feature map.The enhancement improves discriminability and robustness of the global feature and thus improves the recognition accuracy.In the channel dimension,this dissertation proposes a Channel Convolutional Deep Neural Network as the backbone network of the global feature enhancement model.The channel convolution operation is utilized in the network to model pixel-wise channel correlation of global features,which increases the discriminability of global features.In the spatial dimension,a single-stage multi-attention module based on Long and Short-Term Memory network is proposed.This module uses Long and Short-Term Memory network to guide the spatial attention machine and achieves multi-target attention on feature map.As a local feature extractor,this module can supplement the image details of the global feature extracted by the backbone network and improve the discriminability and robustness of the global feature.To verify the effectiveness of the global feature enhancement model,Rank-1 accuracy and mean Average Precision are tested on three public person re-identification datasets.Experiments show that the global feature enhancement model can extract more discriminative image representations compared with other feature extraction algorithms.In addition,the results of comparative experiments and ablation experiments demostrate that the enhancement of global features both from the spatial demension and the channel dimension can improve the identification accuracy.