Patch-Wise Blind Image Deblurring via Michelson Channel Prior

Motion blur exists in many computer vision tasks, including faces, texts, and low-illumination images etc. It has been proved that Dark Channel Prior (DCP) and Bright Channel Prior (BCP) can both help the image deblurring by enhancing the dark or bright channel pixels. However, the pixels between the dark channel pixels and the bright channel pixels are not taken into consideration, which limits the deblurring performance. A novel image channel is proposed in combination with dark channel and bright channel in this paper to consider the effects of the all types of pixels, namely, Michelson channel pixels. Secondly, as the image channels are built based on the series of image patches with different blur kernels, a new method is developed to estimate the blur kernel and can measure the similarity between neighbored kernels. Meanwhile, to perform accurate kernel estimation, the L0 regularization is applied into the algorithm framework. In the process of image deblurring, by enhancing the Michelson channels and retaining the other channels of the image, we can capture sharper image detail and eliminate the ringing artifacts of the recovered images. Massive experimental results demonstrate that the proposed method is more robust and outperforms the existing art-of-the-state of unsupervised image deblurring methods on both synthesized and natural images.