Abstract:
The traditional pedestrian re-recognition method uses manual feature extraction, which has high cost and is difficult to be applied to recognition tasks in complex scenes. The application of deep learning to pedestrian re-recognition can make the model have the ability to extract features independently, and the recognition effect is significantly improved while the cost is reduced. The deeper network can improve the feature expression ability of the network, but the gradient will disappear with the increase of network layers. Residual network can alleviate the problem of gradient disappearance, but the extracted feature information is difficult to be used reasonably. In this paper, the residual network is optimized and the coordinate attention mechanism module is introduced. The coordinate attention mechanism module is used to strengthen the feature information of high contribution rate and weaken the feature information of low contribution rate to improve the network feature expression ability. Another important factor affecting the recognition effect of pedestrian re-recognition model is the phenomenon of occlusion in part of pedestrian images. In this paper, the data enhancement method of grid mask is introduced to reduce network overfitting and improve network generalization ability, which effectively alleviates the problem of occlusion in pedestrian images in real scenes. Finally, difficult triplet loss is used to supervise and train the network. The experimental results show that the rank-1 value of this algorithm can reach 78.7%, 75.8%, 95.7% and 89.6% on CUHK03-Label, CUHK03-Detect, Market-1501 and DukeMTMC-reID datasets, respectively. The mAP values were 78.7%, 76.3%, 73.1% and 88.2%, respectively.