Edge Enhancing Based Blind Kernel Estimation for Deep Image Deblurring

Blind image deblurring problem needs to estimate both the latent image and the blur kernel, which is an active and challenging task in image processing and computer vision. To tackle this ill-posed problem, various regularization techniques have been developed. In this work, we propose an edge-enhancing regularization-based blind kernel estimation model. Our work is inspired by the interesting observation that the gradient amplitude over the real edge will diminish after blurring process while it will increase in the smooth region near the real edge. The novelty is the gradient-based regularity, which is highly effective in recovering image edges, and facilitating kernel estimation. To solve the proposed model, we present an alternating iterative algorithm, which shows good numerical convergence. After estimating the blur kernel, we adopt existing excellent deep learning-based non-blind image deblurring techniques to obtain the final deblurring results. Experiments on extensive datasets show that our method can achieve the state-of-the-art performance among the related blind kernel estimation methods.