Image Super-Resolution Based on Residual Block Dense Connection

Abstract

Image super-resolution models based on convolution neural networks are facing problems such as gradient disappearance, gradient explosion, and insufficient feature utilization. This paper proposes an image super-resolution model based on feature fusion of dense connection of residual blocks. The key contributions are as follows: (1) residual block mechanism, which can make full use of the hierarchical features extracted from the residual block to alleviate the shallow feature losing. (2) In order to extract more representative key features, the feature of each level extracted from residual blocks is input into subsequent residual blocks by dense connection mechanism. (3) local feature fusion is used in a single residual block, and global feature fusion is used in the tail of the model, so that the shallow key information can be transferred to the reconstruction layer as much as possible. Empirical experiment is deployed on four benchmark test sets (Set5, Set14, Urban100 and BSDS100), the results show that both the peak signal-to-noise ratio and structural similarity are improved. (Source code: https://github.com/brown-cats/SR_RFB).