Image Pedestrian Re-identification Based on ShuffleNet-Beyong and Batch-DropBlock

Due to the changes of posture, occlusion and illumination intensity in pedestrian images, the accuracy of pedestrian re- recognition will be reduced. Researchers have used a large number of methods and experiments to improve the accuracy of recognition, but they did not consider that while improving the accuracy, they should reduce the amount of calculation as much as possible. In the current internet era, mobile terminal devices are very popular, so the neural network with large amount of computing is not suitable for these terminal devices. In order to solve this problem, this paper first proposes a ShuffleNet-Beyong unit, which add two 1*1 convolution cores to the left branch of ShuffleNet unit to extract more robust features, and add channel shuffle after the last 1*1 Gconv of the right branch of ShuffleNet unit to solve the problem of channel information non circulation. Then, a ShuffleNet-Beyong network is constructed based on these ShuffleNet units, and an image pedestrian re-recognition network structure based on the ShuffleNet-Beyong network and the Batch-Dropblock network is proposed. In this network structure, the ShuffleNet-Beyong network is used as the backbone network to extract features. After the backbone network, the Batch-Dropblock network is used to extract local features, which are fused with the global features to obtain the final feature expression, so as to obtain more robust features. The comparison of experimental results on Marketing-1501 and DukeMTMC-reID datasets show that the image pedestrian re-recognition network structure proposed in this paper is very robust, which can reduce the amount of calculation and improve the accuracy of recognition, and can achieve the same effect in the process of ablation experiments.