Research Of Person Re-identification And Pruning Of Deep Learning Models

In recent years,countries and individuals have paid more and more attention to security issues,security surveillance has become more and more common,and it has become more and more difficult to find information from massive videos.The purpose of the person reidentification is to establish a correspondence between the images of the same pedestrians taken by different cameras through a computer,and person rei-dentification plays an important role in the field of intelligent security.In recent years,methods based on deep learning techniques have made significant progress in the fields of computer vision,speech,and natural language processing.For the person reidenti-fication,the deep learning based method has become the mainstream method.But the existing deep learning method has insufficient accuracy.The existing methods rarely consider the different granularity details of the person entity.Also not to make use of the relationship between features of different granularities.Not only that,deep learning-based approaches now have high demands on both compute and storage and cannot be deployed on resource-constrained embedded devices.In this paper,the deep learning method for person reidentification and the pruning of deep learning model are studied.The main research contents are as follows:1.Multi-granularities Feature Boosting Network for person reidentification is pro-posed.Different from the existing person reidentification algorithm,the network intro-duces the attention mechanism and the feature boosting module to combine the differ-ent granularity details of the person entity.The multi-task learning method trains the network.Good experimental results have been achieved on major datasets,especially on the CUHK03 dataset using the new test protocol,the accuracy of rank-1 and mAP reached 79.1 and 77.1.2.A convolution filtering pruning method combining L2 paradigm and Taylor first-order expansion is proposed.The method combines the L2 paradigm and the Taylor first-order expansion of two pruning standards.The Multi-granularity Feature Boosting Network is improved by 2.5 times in the compression ratio and 4.2 times the acceleration ratio after the pruning,and the loss of the accuracy is acceptable.