Research On Person Re-identification Methods Via Multi-task Adversarial Learning

The extensive deployment of surveillance cameras in hospitals,schools,shopping malls and other public places,not only ensures public safety and potential troubleshooting accidents but also brings great demand for videos and images analysis.Traditional monitoring methods based on manual processing and extraction have the disadvantages of high error rate,slow speed and inefficiency.Thus,intelligent video monitoring system based on deep learning came into being.As a vital technology of intelligent video surveillance data processing,person re-identification（Re ID）leverages computer vision technology to judge a specific person in images or video sequences obtained by different cameras based on person detection.It is usually regarded as a sub-problem of person retrieval with great research significance and broad application prospects.In the actual scene,person Re ID technology still possesses the following challenges.Firstly,multiple cameras featuring 24-hour and all-weather capabilities collect massive videos and images,however,there will be limited training samples of some pedestrians.Secondly,pose variations caused by the different viewpoints of cameras and person movement will lose pedestrian information and cause appearance changes greater than identity differences,which leads to existing works learning robust person features hardly.Thirdly,to improve the identification accuracy,existing Re ID works usually adopt the complex network,making it difficult to deploy to business scenarios with high real-time requirements.To solve the above problems,this thesis has carried out the following innovative works on diversity sample generation,similar person feature learning and lightweight network design.（1）Diversity person image synthesis method.Aiming at the problem of poor diversity of specific pedestrian samples in a big data environment,we propose a conditional variational generative adversarial network（CVGAN）based on variational inference.By using the pose estimator and designing a variational inference-based encoder,CVGAN decouples both pose and appearance features of the same person and realizes image generation with the arbitrary pose.Experiments on three benchmarks demonstrate that CVGAN achieves superior performance to most mainstream methods.（2）High-resolution person image synthesis method.Existing advances manipulating person image synthesis lack texture details for varying poses or appearances,this thesis presents a person image synthesis Siamese generative adversarial network（PS²GAN）.Image generative network embedded in Siamese is designed to re-synthesize image with the target pose,and a novel contrastive-pose loss further regularizes the generative process.Additionally,a nearest-neighbor loss computes the difference between fake and real images to make high-level information closer.This method has achieved good results on multiple public datasets.Using the generated samples to expand the Re ID training set boosts the recognition performance.（3）Multi-task pose-unrelated feature learning method.Siamese GAN with variational inference based on reinforcement learning（RL-VGAN）is devised for alleviating the pose variations disturbed by similar pedestrians.The proposed variational generative network is used for sample generation and similar sample adversarial learning in Siamese network.Combining the advantages of deep learning and reinforcement learning decision-making improves the ability of the variational generation network to generate diverse samples and the robustness of the RL-VGAN to similar sample interference.A three-stage training strategy is used to realize image generation and person Re ID in the training stage,which handles the difficult identification of pedestrian identity characteristics caused by pose changes and similar pedestrian interference.This method consistently achieves better identification performance than its counterparts on multiple datasets.（4）Filter pruning based on evolutionary algorithms.Aiming at the problems of many training parameters and a large amount of calculation in the person Re ID model,a filter pruning method based on evolutionary algorithms（EAFPruner）is proposed.Firstly,the network pruning process is modeled as a combinatorial optimization problem solved by an evolutionary algorithm,and uses L₁ norm as the redundancy criterion to evaluate the importance of filters.In addition,the adaptive batch normalization is adopted to quickly evaluate the strong relationship between the performance of various pruned networks and their fine-tuned accuracy to reduce the fine-tuning time consumption.Finally,the pruned network with the highest precision is chosen to fine-tune and evaluate the Re ID performance.The experimental results are carried out on four-person Re ID datasets and verify the effectiveness of the EAFPruner.（5）Based on above four works,a person Re ID verification system is designed via multi-task adversarial learning.This system has two major parts:parameter setting sub-module and multi-task adversarial learning person Re ID sub-module,which visually displays the parameter setting of the model and the training process of the network.This dissertation has 64 figures,21 tables and 189 references.