Lightweight dynamic attention network for single thermal image super-resolution

The embedding of attention mechanism in convolutional neural networks (CNN) effectively improves the performance of single image super-resolution (SISR). However, consistent employ of attention modules at distinct depths of the CNN failed conduct the congruous gain, or even degrades performance. In this paper, we propose LDANet, a lightweight SISR network based on dynamic attention mechanism for thermal image. The dynamic attention blocks in LDANet dynamically rescale the attention and non-attention branches according to input features. Specifically, the attention branch composed of pixel- and channel-wise attention blocks to extract the most informative features in pixel domain and channel dimension, respectively. While the no-attention branch consisting of single convolutional layer for extracting features that are ignored by the attention branch. Innovatively, we adaptively and averagely weight the average pooled and standard deviation pooled features within the channel attention block to fully take advantage of the pooled features. Quantitative and qualitative experiments on three thermal image testing datasets with ×\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\times $$\end{document}2, ×\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\times $$\end{document}3 and ×\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\times $$\end{document}4 scale factors and plentiful scenes show that, compared with SISR models of similar size scope, the proposed LDANet accomplishes superior high-resolution thermal image reconstruction performance.