Research And Implementation Of Super-resolution Algorithm For Vehicle Rear-photographed Images In Traffic Surveillance Scenes

Low-resolution images are often accompanied by blurry textures and unclear details,while high-resolution images contain richer information and better perceived quality.Therefore,the goal of super-resolution reconstruction technology is to generate a corresponding high-resolution image based on the input low-resolution image,and add detailed information while maintaining the structural features of the original image.With the accelerated construction of smart cities,the video surveillance of public transportation plays an important role in improving public safety.In super-resolution tasks,one of the research bottlenecks is that it is difficult to obtain the low-high resolution data pairs.The current mainstream super-resolution methods focus on reconstructing degraded images using artificially fixed degraded cores.The effect of low-resolution images is more limited.Therefore,it is a very challenging and practical research task to carry out super-resolution reconstruction of the rear-shot image of the vehicle in the real traffic surveillance scene.In order to obtain low-high resolution images pairs under traffic surveillance scene,this thesis divides the task into three steps.Firstly,the down-sampling network is adopted to learn the degradation process of images from high to low resolution under this particular scene.Then the learned model is used to generate pairs of low-high resolution data.Finally,supervised training is carried out based on pairs of low-high resolution images.Based on this training strategy,we first propose a down-sampling network based on residual structure and generative adversarial network,which down-samples the high-resolution images to low-resolution images while retaining the original scene features.Secondly,we propose a super-resolution network and integrate the residual aggregation module.We design the loss functions based on subjective evaluation metric to improve the perception quality of reconstructed super-resolution images.In addition,we construct a vehicle post-shot image data set,which can facilitate the research of vehicle super-resolution algorithms and the evaluation of objective and subjective metrics.In order to evaluate the effectiveness of the down-sampling network and the super-resolution network,a number of experiments are evaluated on the Vehicle6K dataset.The results show that the quality of the low-resolution image generated by the down-sampling network is better than the existing methods.The proposed super resolution reconstruction algorithm is superior to the existing unpaired super resolution method in both quantitative and qualitative evaluation.