Deep Convolutional Neural Network with Feature Fusion for Image Super-Resolution

Abstract

Recently, deep convolutional neural networks (CNNs) in single image super-resolution (SISR) have received excellent performance. However, most deep-learning-based methods do not make full use of low-level features extracted from the original low-resolution (LR) image, which may reduce the quality of reconstructed image. To address these issues, we propose a method which can connect the low-level features from almost all convolutional layers. Our method use the interpolated low-resolution image as input, employ many skip-connections to combine low-level image features with the final reconstruction process, these feature fusion strategies are based on pixel-level summation operations. After merging the previous convolution features, residual images are used to directly reconstruct high-resolution (HR) images. Experiments demonstrate that the proposed method is superior to the state-of-the-art methods.