Deep CNN for Single Image Super Resolution Using Skip Connections

Abstract

With the current progress in deep learning domain, quite a few deep learning based models have been developed to address the challenges of Single Image Super Resolution (SISR) task. Recent trend in SISR models is towards increasing the depth of the model to achieve a high accuracy through multiple contexts learned at various depth of the network. Skip connections between different layers of the deep network help in utilising multi-context features and also help to address the issue of vanishing gradients but at the cost of increased computational complexity. In this paper, a deep CNN with a minimum number of skip connections is devised to derive a High-Resolution (HR) image from a Low-Resolution (LR) image. A detailed experimental analysis is carried out on skip connection parameters for a given depth to optimise network parameters. The proposed model gives better results compared to existing methods for multiple scales on benchmark data-sets.