Local spatial information for image super-resolution

Abstract

Image Super resolution plays a crucial role in many applications, such as medical imaging, remote sensing, and security surveillance. Recently convolutional neural network are becoming mainstream in computer vision. Most CNN based super resolution methods cannot fully exploit the entire feature from the original image, and thus the corresponding results will appear low resolution. In this paper, we propose a new network which can reconstruct a high resolution images by upscaling the low resolution images layer by layer with a small scale factor. This strategy helps network to possibly avoid of losing information. The existing CNN models involved bicubic interpolation for preprocessing, which leads to large feature maps and high computational loads. To settle of this problem, the proposed network directly extracts features from the input images, without using preprocessing. In addition, the proposed network investigates the spatial information which is represented by dissimilarities between a low resolution image and its corresponding high resolution by adopting a global residual learning. This differentiable strategy is inserted into the proposed network, to dynamically extract the feature maps. The proposed model not only achieves a compatible performance with the existing prominent methods but also, efficiently reduce the computational expenses.