Image Deblurring based on a U-shaped Network for Vehicle Surveillance Scenarios

Motion blur is a common problem in the field of surveillance scenarios, and obstructs the acquisition of valuable information. With the development of deep learning (DL) technologies, a series of CNN-based methods have been developed for image deblurring and achieved significant improvement. However, most existing deblurring methods aim to handle the global blur while in practical local blur is more common. For example, the motion blur is caused by local moving vehicles and the performances of the current methods are poor in the vehicle surveillance scenarios. Based on this observation, this paper proposes a deblurring method, called DeVeh, for the vehicle surveillance scenarios. The architecture of the proposed method is mainly based on a U-shaped network (UNet), and applies an Intersection over Patch (IoP) factor and a supervised morphological loss to improve the performance on the local blur. To verify the effectiveness, a moving vehicle dataset is established over real surveillance scenarios. Comparative and ablation experimental analysis are carried out over it. As a result, our proposal achieves better performance compared to the existing benchmarks.