Image super-resolution employing a spatial adaptive prior model

Super-resolution (SR) methods based on total variation (TV) prior model is a very popular method because of its ability of edge preservation. However, as TV prior model favors a piecewise constant solution, some pseudoedges in the smooth regions which are also called block effect may be produced, especially at high noise levels. In order to overcome such shortcoming, an adaptive SR model based on a new edge indicator is proposed. In our proposed model, a robust trilateral structure tensor by simultaneously considering the spatial similarity, gray similarity and gradient similarity is first constructed to examine the image local pattern, and then we develop a new edge indicator based on the eigenvalues of the trilateral structure tensor to identify the local spatial property of each pixel. Finally, an adaptive prior model controlled by the new edge indicator is proposed, in which the prior model at edges is approximate to the L1 norm in order to preserve edges, while the prior model in smooth regions and noises is approximate to the L2 norm in order to remove the noises. Experimental results on both simulated and real image sequences, our proposed method can better preserve edges and reduce the block effects in smooth regions when compared with the state-of-the-art methods.