Single Image Super-resolution Based on Residual Learning

Patch-based learning methods, such as sparse coding, are by far one of the most dominant ways to handle the single image super-resolution (SISR) issue. However, due to the great success of deep learning, several advanced models based on deep neural networks have been proposed for SISR more recently, gradually revealing its superiority over other counterparts. Therefore, in this paper, we carry on this promising line of work and propose a well-designed network mainly on the basis of residual learning. The key idea of our model is to extract the mean part from the input first in order to lower the impact of background and obtaining two individual components from it. Then, residual learning is applied to mapping the remainder of the input to target high-resolution space, while the mean part is quickly connected to the final output via identity shortcuts. Consequently, our final model conceptually integrates all the above procedures into a completely end-to-end trainable deep network. Thorough experimental results indicate that the proposed method can perform effectively, and is superior to many recently published baselines in terms of both visual fidelity and objective evaluation.