Abstract: In order to solve the ill-posing problem and poor effect of fixed regularization parameter in super-resolution image reconstruction, an adaptive regularization combining the study of sparse representation is proposed. By additional restrictions for compatibility of adjacent patches, a new L_1/2 non-convex optimization model is built. Reweighted L₂ Norm rather than L_p (0<p<1) Norm is applied into the adaptive algorithm for adjustment of regularization parameter. With the help of joint dictionary training method, some important features for improving the quality of reconstructed image are obtained. Experimental results show that the method has significant advantages in denoising and preserving edge details. It is showed that the proposed method not only makes the desired high-resolution images visually clearer, but it also outperforms some traditional methods in both the value of peak signal to noise ratio and structural similarity.

Key words: L_1/2 non-convex optimization, sparse representation, adaptive regularization, super-resolution reconstruction, compatibility of adjacent patches, jointing dictionary