Improved RDN image super-resolution reconstruction network based on multi-scale feature fusion

Abstract

An improved RDN image super-resolution reconstruction network incorporating multi-scale residual dense network (MSRDN) was proposed to address the problem of inadequate use of image features in common single-image super-resolution (SISR) reconstruction networks. Firstly, the shallow features were extracted from the input low-resolution image. Then, the feature extraction module was jointly constructed using convolutional layers and local residual learning structures, and the multiplexed structure of the module at different scales was used to fully extract the multi-scale detail features of the image. Next, a top-down and bottom-up feature fusion module was constructed to fully fuse and correlate the collected multi-scale features to construct image features with richer detail information. Finally, the extracted features were sent to the image reconstruction module for the final super-resolution image reconstruction. The experimental results show that the proposed MSRDN network exhibits better visual results compared to networks such as SMSR on the three super-resolution benchmark sets of Set5, Set14 and BSD100, with the peak signal to noise ratio (PSNR) the obtained peak singal improved by 0.8 dB on average, and structural similarity (SSIM) by an average of 0.02.

Keywords

• super-resolution
• multi-scale features learning
• feature fusion
• rdn network