Efficient Blind Image De-Blurring using Dark Channel Prior

Abstract

Image de-blurring is classified into blind and non-blind image de-blurring. In nonblind image de-blurring, blurring function is known during the extraction of the true image from the degraded one. In blind image de-blurring, blurring function and true image are unknown. Therefore, blind image de-blurring is double ill posed problem as there is no unique solution to blur kernel and restored image. Dark channel constitutes of minimum intensity pixels and prior is selection of minimum intensity pixels. Existing methods make assumptions on blur kernels, latent images or both. Numerous methods assume sparsity of image gradients, which has been widely used in low-level vision tasks including de-noising, stereo and optical flow. Secondly, De-blurring methods formulated within maximum a posterior framework favors blurry images over clear images so they use heuristic selection method which increases computational complexity. State of the art image de-blurring techniques perform better for specific images like face, text or low illumination. To overcome above problems blind image de-blurring using dark channel prior is proposed due to its computational efficiency. The proposed technique employs discrete wavelet transform to improve the efficiency and regularization to minimize the dark channel of the recovered image. The proposed method favors clean images over blurred images in the restoration process. The norm is highly non-convex and the optimization involves a non-linear minimum operation. An approximate linear operator based on look-up tables is used for the min operator and to solve the linearized minimization problem by halfquadratic splitting methods. The algorithm does not require heuristic edge selection steps or any complex processing techniques in kernel estimation. The proposed algorithm converges quickly in practice and can be extended to non-uniform deblurring tasks. Experiments are conducted on face, text and low illumination blur image data sets to evaluate the performance of proposed method. Experimental results show that dark channel prior based image de-blurring is effective for uniform and non-uniform deblurring. The proposed algorithm performs better on de-blurring natural images, and performs favorably against specialized methods for faces, texts, and low illumination conditions.