Biblio

This paper aims at building a sparse deep autoencoder network with joint regularizations for image super-resolution. A map is learned from the low-resolution feature space to high-resolution feature space. In the training stage, two autoencoder networks are built for image representation for low resolution images and their high resolution counterparts, respectively. A neural network is constructed to learn a map between the features of low resolution images and high resolution images. Furthermore, due to the local smoothness and the redundancy of an image, the joint variation regularizations are unified with the coupled autoencoder network (CAN). For the local smoothness, steerable kernel variation regularization is designed. For redundancy, non-local variation regularization is designed. The joint regularizations improve the quality of the super resolution image. Experimental results on Set5 demonstrate the effectiveness of our proposed method.