Image super-resolution network based on multi-scale adaptive attention

被引：1
|
作者：
Zhou Y. [1 ,2 ]
Pei S. [1 ]
Chen H. [1 ,2 ]
Xu S. [1 ]
机构：
[1] School of Artificial Intelligence, Hebei University of Technology, Tianjin
[2] China Hebei Control Engineering Research Center, Tianjin
来源：
Guangxue Jingmi Gongcheng/Optics and Precision Engineering | 2024年 / 32卷 / 06期
关键词：
adaptive weights; attention mechanism; multi-scale feature; progressive information interaction; super-resolution;
D O I：
10.37188/OPE.20243206.0843
中图分类号：
学科分类号：
摘要：
Aiming at the problem that most image super-resolution methods cannot fully extract features by using single-scale convolution，an image super-resolution network based on multi-scale adaptive atten⁃ tion is proposed. To fully use the contextual information in each hierarchical feature，a multi-scale feature fusion block was designed，whose basic unit consists of an adaptive dual-scale block，a multi-path progres⁃ sive interactive block，and an adaptive dual-dimensional attention sequentially in series. Firstly，the adap⁃ tive dual-scale block autonomously fused the features of two scales to obtain richer contextual features；sec⁃ ondly，the multi-path progressive interactive block interacted the output of the adaptive dual-scale block in a progressive way to improve the correlation between the contextual features；lastly，the adaptive dual-di⁃ mensional attention autonomously selected different dimensions of the attention to refine the output fea⁃ tures，which makes the output features more discriminative. The experimental results show that on Set5，Set14，BSD100 and Urban100 test sets，the method of this paper improves the PSNR and SSIM quantita⁃ tive metrics compared to other mainstream methods，especially for the Urban100 test set，where texture details are difficult to be recovered，the method of this paper improves PSNR and SSIM metrics by 0. 05 dB and 0. 004 5 respectively compared to the existing optimal method，SwinIR，with the scaling fac⁃ tor of ×4；in terms of visual effect，the reconstructed images in this paper have more texture details. © 2024 Chinese Academy of Sciences. All rights reserved.
引用
收藏
页码：843 / 856
页数：13
相关论文
共 33 条
  • [1] HIJJI M, KHAN A, ALWAKEEL M M，, Et al., In⁃ telligent image super-resolution for vehicle license plate in surveillance applications［J］, Mathematics, 11, 4, (2023)
  • [2] CHEN S L, ZHAO C B，, Et al., Large-scale individual building extraction from open-source satellite imagery via super-resolution-based instance segmentation approach［J］, ISPRS Journal of Photogrammetry and Remote Sensing, 195, pp. 129-152, (2023)
  • [3] SUN Y W，, Et al., An arbitrary scale super-resolution approach for 3D MR images via implicit neural representation［J］, IEEE Journal of Biomedical and Health Informatics, 27, 2, pp. 1004-1015, (2023)
  • [4] Image super-resolution using deep convolutional networks［J］, IEEE Transactions on Pattern Analysis and Ma⁃ chine Intelligence, 38, 2, pp. 295-307, (2016)
  • [5] CABALLERO J, Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition （CVPR）, pp. 1874-1883, (2016)
  • [6] KIM J, LEE K M., Accurate image super-resolution using very deep convolutional networks［C］, Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition （CVPR）, pp. 1646-1654, (2016)
  • [7] TAI Y, LIU X M., Image super-resolution via deep recursive residual network［C］, 2017 IEEE Conference on Computer Vision and Pattern Recognition（CVPR）, pp. 2790-2798, (2017)
  • [8] KIM H，, Et al., Enhanced deep residual networks for single image super-resolution ［C］, 2017 IEEE Conference on Computer Vision and Pattern Recognition Workshops （CVPRW）, pp. 1132-1140, (2017)
  • [9] CHENG D Q, ZHAO J M, Et al., Multi-scale dense feature fusion network for image super-resolution［J］, Opt. Precision Eng, 30, 20, pp. 2489-2500, (2022)
  • [10] TDPN：texture and detail-preserving network for single image super-resolution［J］, IEEE Transactions on Image Processing：a Publication of the IEEE Signal Processing Society, 31, pp. 2375-2389, (2022)
1234