In order to reconstruct more boundary information for the weak-boundary retina image via sparse representation, the paper proposes a super-resolution (SR) algorithm by combining edge difference with semi-coupled dictionary learning (ED-SCDL). Firstly, we consider the edge difference between low-resolution (LR) image and degraded version of reconstructed image as constraint term of the proposed ED-SCDL model to preserve the edge information. Then, in order to improve the contribution of edge difference constraint on preserving boundary information, the adaptive regular parameter is explored via the approximate Laplacian distribution of edge difference. Experiments on retina images demonstrate that our model is superior to other methods, especially retina images with weak edges.