Study On Person Re-identification Based On Deep Neural Network

Nowadays social security is getting more and more attention,cameras as an important tool for maintaining public security are around us.Person re-identification is a research technology based on the development of cameras;it identifies the same person in the pictures taken by different cameras.Person re-identification is mainly used for criminal investigation,looking for lost elderly people and children,analyzing people’s shopping behavior in shopping malls.The work of this thesis abandons the manual process of searching person in massive monitoring video,using an end-to-end deep learning method,so that we could save a lot of time.This thesis makes the following three improvements on deep neural networks by studying some classical person reidentification techniques:(1)In the data processing phase,a new data augmentation technique is applied.Different from the traditional data augmentation techniques such as cropping and panning in the image,this thesis adopts the method of random occlusion in the image,whether the position in the image or the size of the occlusion position is random.In this process,training images with various levels of occlusion are generated,which reduce the risk of network overfitting and make the model robust to occlusion.(2)In the network training phase,this thesis decorrelated the first fully connected layer.The correlation of the weights of the fully connected layer has an impact on the performance of the model.In this thesis,the singular value matrix decomposition is performed on these weight matrices,then a left-unitary matrix and a singular value matrix are used for taking place of the original matrix,but this operation cannot directly improve the performance of network.This thesis also adopts a restraint and relaxation iteration training scheme,which train the network fixing the orthogonal layer parameters in first step and train the whole network in second step.(3)In the testing phase of the network,this thesis uses a k-reciprocal re-ranking method.The initial ranking based on Euclidean distance often contains mismatches.This thesis adopts a stricter re-ranking method based on the initial ranking——two pictures must be in the top-k ranked images by each other.In order to facilitate the calculation,this thesis encodes the k-reciprocal feature into a single vector and assigns different pictures with different weight values.Our method has been experimented on Market-1501 and CUHK03,it achieves a great performance,the results show that,rank-1 accuracy is improved to 86.75% and mAP is improved to 80.44% on Market-1501 for ResNet-50,while on CUHK03,rank-1 accuracy is improved to 49.23% and mAP is improved to 50.54%.