Blur Kernel Optimization: A New Approach to Patch Selection with Adaptive Kernel Estimation

Recently, many effective approaches appeared in the field of blind image deconvolution to reduce the computational cost. Using multiple smaller regions instead of whole image not only make the restoration efficient but also improves the results by discarding the ineffectual regions. It is observed that a study is needed to compare different methods for the selection of useful image patches and different schemes to utilize their blur kernels, which is aimed in the present work. A new patch selection method using Contrast based blur invariant features (CBIF) is proposed to find the useful regions which gives better results compared with others e.g., speed-up robust features (SURF), local binary patterns (LBP), local phase quantization (LPQ), maximally stable extremal regions (MSER), Canny and Sobel. In addition, gradually increasing contrast stretched levels shown to give better results compared with commonly used multiscale framework to avoid false local minima. It is also proposed that blur metric by Crete applied on latent image can be used for the selection of better kernel. It is observed that an effective strategy can give good results even when the patches are not selected carefully. The best results are obtained when our proposed patch selection is used with our “selective kernels averaging” scheme.