Single-image super-resolution via joint statistic models-guided deep auto-encoder network

Abstract

Recent researches on super-resolution (SR) with deep learning networks have achieved amazing results. However, most of the existing studies neglect the internal distinctiveness of an image and the output of most methods tends to be of blurring, smoothness and implausibility. In this paper, we proposed a unified model which combines the deep model with the image restoration model for single-image SR. This model can not only reconstruct the SR image, but also keep the distinct fine structures for the low-resolution image. Two statistic priors are used to guide the updating of the output of the deep neural network: One is the non-local similarity and the other is the local smoothness. The former is modeled as the non-local total variation regularization, and the latter as the steering kernel regression total variation regularization. For this unified model, a new optimization function is formulated under a regularization framework. To optimize the total variation problem, a novel algorithm based on split Bregman iteration is developed with the theoretical proof of convergence. The experimental results demonstrate that the proposed unified model improves the peak signal-to-noise ratio of the deep SR model. Quantitative and qualitative results on four benchmark datasets show that the proposed model achieves better performance than the deep SR model without regularization terms.