Weak texture information map guided image super-resolution with deep residual networks

Bo Fu, Liyan Wang, Yuechu Wu, Yufeng Wu, Shilin Fu, Yonggong Ren

Abstract: Limited by the poor quality of the camera, transmission bandwidth, excessive compression and other factors, low-resolution images widely exist in our lives. Single image super-resolution method is a kind of image processing task which can obtain high-resolution image from corresponding source low-resolution image. With the development of deep learning technology, a series of deep learning methods have brought crucial improvement for SISR problem. However, we observe that no matter how deep the network structures for SISR are designed, they usually have poor performances on restoration of tiny or irregular details, we call them weak texture information. The main reason for this phenomenon is that weak texture information is not obvious relative to the salient features, so as weak texture information feature is less extracted in the process of neural network feature extraction. To address this problem, we propose a SISR method which owns unique weak texture information prediction module and call it as weak texture information map guided image super-resolution with deep residual networks. In our network structure, two auxiliary sub-networks work together to capture grab details information and predict weak texture information. Then, predicted weak texture information will be fused into main sub-network. Therefore, our network can obtain more irregular details which usually miss in deep learning based methods. Finally, both qualitative experiments and visual effects demonstrate the effectiveness of our proposed algorithm. Specifically, our method can restore more irregular image details.