Research On Key Technologies Of Person Re-identification Based On GAN Optimization Model

Person Re-Identification(Person ReID)technology is gradually playing a huge role in applications such as community security and identity verification.However,in the real world,the accuracy of Person ReID is affected by many factors,such as illumination of environment,camera resolution,pedestrian behavior,and occlusion.In order to improve accuracy and overcome these challenges,the end-to-end model based on deep learning is used to construct the Person ReID framework(referred to as end-to-end Person ReID model),and current research mainly focuses on constructing data sets,feature extraction,classification and retrieval.The end-to-end Person ReID model is favored by researchers because of its simple implementation steps and high recognition accuracy.However,the most significant factor affecting the accuracy of end-to-end Person ReID is the size of the training data set.At present,the mainstream data sets of Person ReID are not large enough.Therefore,how to flexibly expand the data set is an important research direction in this field.At this stage,researchers prefer to use low-cost,unlabeled images to expand the training data set,aiming to improve Person ReID accuracy via semi-supervised learning.Based on this background,in this paper,the Generative Adversarial Net(GAN)is applied to generate the unlabeled images,then the original images and the generated images are put into the end-to-end deep neural network model for training,a new reasonable classification algorithm is proposed to tell the categories of persons at the same time,finally retrieval task is completed,so that the accuracy of end-to-end Person ReID is improved.The main works of this paper are as follows:(1)Expanding the data set using unlabeled images.The labeled images in the original data set are put into the Generative Adversarial Net,different qualities of unlabeled generated images are obtained by adjusting parameters in the network and by applying model ensemble algorithm to enlarge training data set.(2)A classification algorithm based on spectral clustering is proposed.During model training process,the losses for the generated images need to be calculated.In this paper,generated images are clustered into the same category as original images by using spectral clustering algorithm,and the distances between the cluster centroids and original images determine the similarity between the generated images and the original images,then the loss of each generated image in semi-supervised learning according to similarity and the different quantities of generated images is obtained,plus under-fitting and over-fitting degrees in the model are both effectively controlled without using complex traditional feature extraction algorithms to improve the accuracy of end-to-end Person ReID.The focuses of this paper are to use Generative Adversarial Net to generate unlabeled images,to calculate the loss of generated images in model training process,to improve the accuracy of classification on mainstream data sets,and to be superior to current best algorithms.