Edge Map-guided Scale-iterative Image Deblurring

Blind image deblurring aims to remove the blur generated from camera motion or moving objects. Advance of deep end-to-end learning methods showed superiority in removing non-uniform motion blur, but there still exists unsharp blurriness due to the pixel-wise loss in a restored image using deep deblurring methods. Therefore, we propose a simple and effective iterative edge map guidance to restore spatial details. The proposed method extracts the edge map of the blurred input image prior to the image deblurring process and restores it to a clear edge map. We use the edge map information for image deblurring task. Therefore, unlike conventional methods, the proposed method performs deblurring by considering the edge information of the clear image, which leads to restored images with sharper edges. Furthermore, the proposed method iteratively reconstructs sharp images and edge maps on multiple scales. This iterative scheme can further improve performance due to the advantages of weight-sharing with multi-scale training. We found that the proposed method showed 0.26dB higher PSNR compared to the original method for GoPro dataset.