The Algorithm Research Of Motion-blurred Image Restoration Based On Patch Priors

An image will be blurred globally if the camera shakes when taking a picture. How toremove blur and get a clean image has become an important content in image restoration.The difficulty obtaining a sharp image is that the blur kernel, sharp image are bothunknown, making the problem ill-posed. We tackle this problem with patch priors to helpblur kernel estimation.In the image restoration, given a blur kernel we estimate a latent sharp image, andthen use the latent image and the blurred image to estimate a new blur kernel, and repeatabove steps until the change of the latent image is less than a certain threshold. Ouralgorithm builds a image pyramid to use a multi-scale strategy, and estimate a blur kernelfor each level from top to bottom which is used for next level as a initial value. Since theestimated latent image has a direct impact on next blur kernel estimation, so thatalgorithms which can estimate high-quality latent image, will estimate the blur kernelmore accurately. The key to estimate the latent image is to use prior knowledge on latentimage. The prior knowledge of the method is following: image gradients are sparse. Patchvariance of a sharp edge patch is bigger than that of its corresponding blurred patch, edgepatches can be approximated as all possible combinations of rotations and translations onseveral specific types of edge patches (such as step edge, corner edge, bar edge).Thevariance distribution of edge patches in a latent image should be close to the empiricalvariance distribution of edge patches which are from clean images in nature. Based onabove prior knowledge, a latent image with high quality is estimated, and then a moreaccurate blur kernel is estimated.The effectiveness of our method is demonstrated by experimental results on bothsynthetic and real-world examples. Experiments show that our proposed algorithm iscapable of accurately estimating the blur kernels of camera motions. The algorithm alsohas achieved good results for motion-blurred image recovery in low-light conditions..