Research On Person Re-identification Method Based On Region Sparse Attention Network

Person re-identification can be applied to criminal investigation,social security and other fields.Many person re-identification methods have been proposed,which are mainly divided into two categories: person re-identification methods based on artificial design features and person re-identification methods based on deep learning.Early person re-identification generally adopts the method based on manual design features,which is theoretical and requires little computation.However,due to the manual design of such features,the cost is relatively high and person features are easy to be ignored.In recent years,the rapid development of deep learning methods has provided new opportunities for the application of person re-identification.The combination of deep learning and person re-identification can not only avoid a lot of labor costs,but also enable the machine to automatically learn and optimize person characteristics,thus improving the performance of person re-identification algorithm.However,due to the low resolution of person photos and in deep learning,a large amount of available information will be lost in the process of decreasing the size of feature maps.In view of the remarkable achievements of artificial design features and deep learning in person re-identification,based on the above two research directions,this paper proposes a region sparse attention network that can be used for person re-identification.The network can effectively avoid the necessary information loss in convolution by adopting the data augmentation method using random region batch occlusion and embedding the sparse attention mechanism.The main methods are as follows:1)A squeeze and activation network embedded with sparse attention.The network is improved from the squeeze and activation network.Firstly,the squeeze and activation module in the squeeze and activation network is extracted and normalized to generate an attention module: a normalized squeeze and activation module.Then four normalized squeeze and activation modules are applied to the 5 convolution layers of the residual network respectively.Finally,4 shortcut are added between the 5 convolution layers to construct a squeeze and activation network embedded with sparse attention.Experiments show that compared with the squeeze and activation network,the Rank-1 and mAP of the proposed method on the person re-identification dataset Market-1501 are improved by 4.2% and 4.4% respectively.2)Data augmentation method using random region batch occlusion.Firstly,the person photos of a training batch are evenly divided into 6 regions horizontally.Then 2 of the 6 regions are randomly blocked,so that the neural network only needs to process the remaining 4 regions.Finally,the 4 regions are processed by global maximum pooling and classified.Experiments show that compared with the squeeze and activattion network,the Rank-1 and mAP of the proposed method on Market-1501 are improved by 5.1% and 7.8% respectively.Through tests on multiple person re-identification datasets,the results show that the squeeze and activation network embedded with sparse attention has good feature extraction capability,and the random region occlusion of person pictures in batches also enables the network to pay more attention to small features.The combination of the two methods has achieved good results in solving the problems of small target persons,complex person background and similarity of different persons.