Kernel estimation and reconstruction for motion blurred images
图像抖动核函数估计与图像恢复

Motion blur is a common problem in hand held photography, for which this paper presented an efficient and robust method for kernel function estimation and image deblurring. It gave an image blurred by camera motion, first built an image pyramid, and then iteratively estimated the motion kernel and the latent image in a top-down manner. It modeled the motion kernel by a mixture of Gaussian, constrained and restored the latent image with a heavy-tailed distribution of natural images, and predicted the strong edges of the latent image by a shock filter, constraining the gradients and the motion kernel of the image. Moreover, it introduced a hysteresis thresholding method and a re-centering method to suppress the noise of motion kernel and improve the robustness of large motion kernel estimation. At last, it solved the kernel function using the conjugate gradient method with the first order and the second order of image derivatives to reduce the condition number, which consequently accelerated the convergence. Finally, it tested the proposed method on a publically available data set with 32 motion blurred images. As showed in the results, the proposed algorithm restores high quality latent images with clear edges and textures, free from the ringing artifacts and noises.