Reference-based dual-task framework for motion deblurring

Deep learning algorithms have made significant progress for deblurring in dynamic scenes. However, most of the existing image deblurring methods use a single blurry image as the input of the algorithm, which limits the acquisition of information and fails to preserve satisfactory structural texture. In contrast, we present a reference-based dual-task framework to recover a high-quality image by deblurring and enhancing a blurry image under the guidance of a reference image. Specifically, the framework includes two tasks: single image deblurring and reference feature transfer. The single image deblurring task deblurs the blurry image leveraging only the blurry image itself. The reference feature transfer task extracts and transfers abundant textures from the reference image to the coarsely result of the single image deblurring task. Benefiting from the reference image, our proposed method achieves more realistic visual effects with sharper texture details. Experimental results on GoPro, HIDE and RealBlur datasets demonstrate that our method outperforms state-of-the-art methods both quantitatively and qualitatively.