Image Super-resolution Based on Alternate K-Singular Value Decomposition

The coupled dictionary training algorithm in super-resolution based on sparse representation are directly related to the detail recovery capability of the algorithm, but the existing algorithm makes the dictionaries lack texture structure information. This paper proposes an alternate K-singular value decomposition dictionary training algorithm. This algorithm is composed of a training stage and a testing stage. In the training stage, the best low rank approximations of low and high frequency patches are used for the updating of the dictionaries. This method makes the sparse representations of low and high frequency patches becomes more and more similar with the increasing of the iteration number. In the testing stages, the high frequency details can be estimated by multiplying the sparse representations generated with low frequency patches with the high frequency dictionary. The experimental results demonstrate that the proposed algorithm can provide clearer resultant images. Compared with many existing methods, the average peak signal to noise ratio exceeds about 0.3dB and structure similarity exceeds about 0.1.