A New Convolutional Neural Network for Super-Resolution by Global and Local Residual

Abstract

With deep learning applied in image super-resolution (SR), more CNN-based method was applied to image super-resolution problems in recent researches. A lot of training time and memory space were used to make network convergence by tradition SR methods, these defects limit the application of SR. To address this issue, a new convolution neural network structure for single image super-resolution (SISR) was proposed. In this work, network structure designed mainly in three aspects: First, global residual and residual block was combined to make the network convergence faster without batch normalization. Second, residual block was redesigned to make the network easier to train. Last point is that data augmentation was applied to solve the problem of poor network generality when the quantity of train images is insufficient. Experiments proved that our method can restore more high frequency details of images, get better in visual effect, improved the PSNR and SSIM of result.