doi：10.16186/j.cnki.1673-9787.2021110080

Received:2021/11/19

Revised:2022/06/06

Published:2024/01/25

Image super-resolution reconstruction of multi-scale residual dense attention network

NI Shuiping, WANG Shijie, LI Huifang, LI Pengkun

School of Computer Science and Technology，Henan Polytechnic University，Jiaozuo 454000，Henan，China

Abstract: Objective Using a single-scale convolutional network to extract low-resolution（LR） image features could cause a large number of image high-frequency features to be lost.In order to obtain more high-frequency features and reconstruct clearer super-resolution images， Methods a single image super-resolution reconstruction algorithm based on multi-scale residual dense attention network was proposed. Firstly，the convolutional network was used to extract shallow features from low-resolution images and the shallow features were used as input at all levels of the subsequent network.Secondly，the multi-scale residual dense attention blocks at all levels were used to process the image features of the previous network and extracted the high-frequency features of the image.The multi-scale residual dense network was good at extracting richer image features and attention mechanism was fused into the network to make high-frequency region features get more attention.Then，the image features of different depths were extracted at all levels of the network for global feature fusion.Finally，the fused features were up-sampled to output the reconstructed super-resolution image. Results When the upscale factor was set as 4，the network was tested on the SET5，SET14，BSDS100，URBAN100 and MANGA109 datasets，and the peak signal-to-noise ratios were 31.97，28.58，27.57，25.85 and 29.79 dB，respectively.The basic modules in the network were replaced by multi-scale residual dense attention blocks，residual blocks and dense blocks to extract features.The peak signal-to-noise ratio was used as the module performance evaluation standard，and the multi-scale residual dense attention block performed better. Conclusion The network combined with the multi-scale residual dense network could obtain richer high and low frequency information of the image.The attention mechanism was fused to effectively extract the high frequency information in the network，and the super-resolution image with clearer texture could be reconstructed.

Key words:multi-scale residual;dense attention network;super-resolution reconstruction;attention mechanism;high frequency region