DDANet: Dilated Deformable Attention Network for Dynamic Scene Deblurring

Abstract

Image motion blur is a phenomenon that typically occurs due to the movement of dynamic objects or the shaking of a digital camera. These blurring problems have non-uniformity and non-directionality. Recent deblurring research aims to address the blur problems by employing self-attention and scale transformation approaches. The self-attention approach can be affected by the attributes of unrelated images in all spatial domains, and the multi-scale approach incurs high computational costs due to its own recurrent framework. In this paper, we propose a dilated deformable attention network (DDANet) that focuses on relevant blur attributes at all positions and handles significant variations in blur attributes distributed over different spatial domains. DDANet also utilizes a multi-scale architecture to leverage the correlation of blur attributes through progressive spatial variations in the input image. Extensive experimental results on the GoPro benchmarks demonstrate that the proposed DDANet effectively performs blurred image restoration in both subjective and objective evaluations.