Research On Person Re-Identification Based On Deep Learning

With the improvement of people's living standard,person re-identification(Re-id)technology is widely used in intelligent monitoring,shopping guide,human-computer interaction and other fields.Given a monitoring pedestrian image,and computer vision technology is used by Re-id technology to retrieve the pedestrian images taken at different locations and times under other devices to determine whether they belong to the pedestrian image.Currently,the accuracy of person Re-id has been improved to a high level in many photographed pedestrian data sets.However,in the actual scene,due to the influence of shooting Angle,time and place,lighting,equipment resolution and other factors,Re-id technology faces great challenges,so how to improve the accuracy of Re-id still has a long way to go.It has become a trend to apply deep learning technology to Re-id,because compared with traditional feature extraction and distance measurement methods,methods based on deep learning can achieve more robust recognition performance.This paper researches improve performance of Re-id on the Market1501,DukeMTMC-reID and CUHK03 data set,and two improved Re-id methods based on deep learning are proposed.The main research work is as follows:(1)Considering that too few images in the training set will reduce the generalization ability of the training model,the first method firstly expands the data set and adopts the method of deep convolution generative adversarial networks(DCGAN)to train and generate images with similar pedestrian characteristics.Then the generated images and the original data set images are combined for network training and the generated images are normalized to construct the loss function.Finally,the pre-training model based on ResNet50 network is adopted in the training and a Siamese network is constructed to optimize the model by combining identification loss and verification loss.(2)The second method optimizes the method of expanding the data set and refers to other optimization methods.Firstly K-means clustering algorithm is used to cluster and divide the pedestrian data set images,and deep convolution generative adversarial networks is used to generate pedestrian with similar characteristics of cluster images after clustering.According to the characteristics of the generated images,a cluster labeling smoothing normalized loss function is proposed.Then in a new Re-id baseline joined part-based convolutional baseline(PCB)method,two pre-training models based on ResNet50 network and DenseNet12 l network are used for comparison experiment.Experimental results show that the first method used DCGAN to expand data sets can obviously improve the Re-id accuracy,the second approach that expands combined data sets and joins PCB can greatly improve the Re-ID accuracy,and confirmed that the DenseNet121 network is better than the ResNet50 network in the performance of Re-id.