Blur kernel optimization with contrast levels and effectual patch selection using SURF features

Abstract

Single image blind deblurring is a challenging and well known ill-posed problem. Recently, with the emergence of many effective algorithms to estimate blur kernels, the research for blur kernel refinement and for developing fast and reliable methods to utilize effectual image regions is becoming increasingly important. Generally, multiscale framework is used for blur kernel refinement to avoid trapping in local minima, however, we recommend to use images with increasing contrast levels which gradually improves blur kernel estimation. To make the kernel estimation more efficient, we used effectual patches instead of whole image, which not only make the restoration efficient but also improves the results by discarding the ineffectual regions. It is especially well suited for the large satellite images corrupted with atmospheric turbulence, motion blur or objects with uniform background. After extensive analysis and comparison with other methods, speed-up robust features (SURF) based patch selection method is proposed. In addition, masking based on gradient directions is also found useful in suppressing misleading regions. Finally, a new scheme is proposed and analyzed which combine the effectual regions as well as contrast levels. The results are found to be improved significantly using SURF based patches, gradient direction masking and contrast level images. The comparisons show the effectiveness of proposed approach.