An improved image restoration algorithm based on the multi-dictionaries

Sparse representation is a useful and wonderful method in the field of image restoration. Compared with the traditional image restoration algorithm, the image restoration algorithm that is based on the sparse representation can improve the quality of restoration effectively. However, it is difficult to express complicated features of an image with one single dictionary, resulting in discrepancies of massive structure. This paper proposes a novel image restoration algorithm based on the multi-dictionaries that combines image clustering and dictionary training. The broken image is divided into a few blocks and generate an index matrix. And a Steering Kernel Regression Weight (SKRW) algorithm is applied for image clustering. To study and explore the inner structures of the image, blocks are sorted into a few categories based on similar SKRW values. And each category will have its own overcomplete dictionary by training with a K-singular Value Decomposition (K-SVD) algorithm. Compared with using one dictionary to express the whole image, multi-dictionaries are matched with different image structures, therefore the image will have a more accurate sparse representation. The dictionaries and sparse coefficients are updated until achieving convergence and the image can be restored after replacing the patches. Experimental results show that our improved algorithm can restore images damaged by noise or scratches more effective than traditional algorithms.