Edge-Enhanced Deformable Attention Network for Video Deblurring

It is important to recover edge information for improving the quality of video deblurring. This paper proposes an additional edge-enhanced branch to spatio-temporal deformable attention network (STDANet) for video deblurring, named edge-enhanced STDANet (ESTDANet). The combination of Sobel filtered edges and optical flow enable ESTDANet to focus on edges orthogonal to the direction of motion blur, and ESTDANet is expected to sharpen blurred edges. Our experiments demonstrate our proposed method improves the accuracy of STDANet on both synthetic and real blurred datasets.