Research On Pedestrian Re-identification Technology Based On Cycle GAN And Siamese Network

With the development of social science and technology,Person Re-identification(re-ID)technology is widely applied in the intelligent security,criminal investigation,image retrieval,etc.Currently,the re-ID precision has been highly improved in many shot person dataset.But in the actual scene,influenced by shooting angle,occlusion,pedestrian appearance,camera resolution,etc.the re-ID technology still confronts a huge challenge.With the development of deep learning technology,its application in person re-identification becomes a direction in today's research.Compared with traditional distance measurement and artificial feature extraction method,the recognition performance of deep learning method re-ID is more robust.The main research contents of this thesis as follows:(1)Considering that insufficient training data will cause the model to have overfitting problems,firstly,the dataset is expanded,and the Cycle Generated against the Network(Cycle GAN)training is adopted for style transfer to compound the image with person characteristics.Then the generated images combine with the original dataset image for network training.Since many cameras used upon dataset collection have different parameters,the extra noise would generate upon image style migration,and the standard smooth normalized loss function is adopted for the Cycle GAN.Finally,the pre-training model based on convolutional neural network is adopted upon training to achieve the self-similarity and domain-dissimilarity in combination with one Siamese network.(2)The study uses an improved binary triad loss function training method.In view of the fact that too many training iterations of the traditional Siamese network are prone to overfitting and occupy a large amount of computer memory,this paper designs an improved triple loss function.Test results on Market-1501 and other public data sets show that,compared with the traditional Siamese network,this method can greatly reduce the consumption of computer running memory while ensuring a high accuracy.