Image super-resolution via deep residual network

Abstract

Recently, convolution neural network (CNN) has been widely used in single image super-resolution (SR). However, the traditional network structure has the problems of fewer convolution layers and slow convergence speed. In this paper, an image super-resolution method based on deep residual network is proposed. Through the deepening of the network structure, more receptive fields are obtained. Thus, more pixel information is utilized to improve the reconstruction accuracy of the model. The feature extraction process is carried out directly in low resolution space, and the images are sampled by shuffling the pixels at the end of the network. The learning method combining global residual and local residual is used to improve the convergence speed of the network while recovering the high-frequency details of the images. In order to make full use of image feature information, feature maps extracted from different residual blocks are fused. In addition, parametric rectified linear unit (PReLU) is used as the activation function, and the Adam optimization method is used to further improve the reconstruction effect. The experimental results of benchmark datasets show that the proposed method is superior to other methods in subjective visual effects and objective evaluation indicators.