Research On Cross Resolution Person Re-identification Based On Image Super-resolution Technology

As one of the most important technology of automatic video surveillance,person reidentification(Re-ID)has been attracted widespread interest.Most existing Re-ID methods usually assume that all pedestrian images taken from different cameras have uniform resolution.However,in many real-world scenes,due to variations of distance between cameras and persons as well as the deployment setting of cameras,the resolution of person images is usually different.If we directly match pedestrian images with different resolutions,the performance of Re-ID will be affected adversely because of the discrepancy of information amount.To tackle this issue,one potential solution is to combine the super-resolution(SR)technology with the Re-ID method.The main work is as follows:(1)Fast super-resolution convolutional neural network(FSRCNN)only uses specific scale images in the training process,which often ignores the information of other scale images.We propose a cascaded convolutional neural network for image SR(CSRCNN).By cascading three FSRCNN,which each FSRCNN can process a specific scale image.Thus,the network can mine three different scales of image information in the process of network training,making the final restored image achieve better visual effect.Then,we analyze the influence of the SR network on the performance of person Re-ID.We use the framework of SR + Re-ID,Bicubic?SRCNN,FSRCNN?SRGAN?CSRCNN are used as SR modules,Res Net50 are used as feature extraction network,and experiments are carried out on MLR-market1501 person dataset.Experiments show that the introduction of image SR network can restore the lost information of low resolution image to a certain extent.(2)Based on the analysis of the integration compatibility of the SR network and the person Re-ID network,we propose a Multi-scale Deep Feature Representation Based on SuperResolution Generative Adversarial Networks(MFR-GAN)framework for low resolution person Re-ID,which aims to optimize the super-resolution of image and pedestrian matching.We first design three cascaded SR-GANs to increase the resolution of person images with different upscaling factors and then introduce a re-identification network after each SR-GAN to strengthen the representation capability of image features.Experiments on several cross resolution person datasets show that the proposed model can effectively reduce the problem of accuracy degradation in person Re-ID.