Sparse representation based image super resolution reconstruction

The present paper addresses a single image super resolution reconstruction approach based on sparse representation of image patches. The proposed reconstruction process enforces a better sparsity solution which is guided by the sparse prior from the L1 norm optimization process. In the optimization process, an efficient feature extraction operator is used to ensure accurate prediction of the high resolution image patch. The normalized cross correlation is used as a similarity constraint to control the matching of image patch in the sparse framework. Finally, the reconstruction process is made robust to noise by selecting an optimal adaptive sparsity regularization parameter using particle swarm optimization method. In the present work, coupled dictionary training is used to learn the dictionaries. The efficiency of the proposed work is validated with different real and synthetic images. Various image quality metrics demonstrates the superiority of the proposed work over other existing super resolution reconstruction methods.