Image super-resolution using deep belief networks

In this paper, we aim at using Deep Belief Networks (DBNs) to solve the problem of image super-resolution (SR). We exploit the hierarchical structure of the DBNs to capture the non-linear mapping from low-resolution (LR) patches to their high-resolution (HR) counterpart. When a query LR image is input, we divide it into a list of patches, then we put each patch into a forward propagation network which is a trained deep belief network. The output is the predicted HR patches. Finally, we combine the HR patches into expected HR images. We evaluate our approach on a popular dataset which is used in other super-resolution literature. Experimental results demonstrate the performance of our method is superior to several state-of-the-art super-resolution methods both quantitatively and perceptually.