Image Super-Resolution Based on Residual Local Feature Fusion

ABSTRACT In recent years, single image super-resolution (SISR) based on deep neural networks has achieved excellent performance. The emergence of super-resolution generative adversarial network (SRGAN) [1] has also become a pioneering work in the field of Single Image Super-Resolution. However, how to recover texture details is still a challenge in the field of super-resolution. Many current convolutional neural networks fail to make full use of shallow feature information, have high computational complexity, and have difficulty capturing the dependencies among feature channels. So we propose a new generative adversarial network to further improve the visual quality. Firstly, we design a feature-fusion networks (FFN) for use in the generator to fuse feature information at different scales; secondly, we introduce a residual local feature networks (RLFN) into the FFN to make the network easier to train and converge. In addition, we introduce contrastive perception channel attention (CCA) to improve the cohesion between channels. The experimental results show that our method achieves better performance in both visual effect and quantitative evaluation metrics.