Thermal image super-resolution via multi-path residual attention network

Convolutional Neural Networks (CNN)-based Single-Image Super-Resolution (SISR) methods for RGB images have flourished rapidly. However, thermal images SR methods based on CNN are rarely studied. The performance of existing deep SR methods is limited by the narrow receptive field of single small convolution kernel (e.g., 3 x 3). In this paper, we propose a thermal image SISR deep network MPRANet, combining multi-path residual and attention blocks. Specifically, an innovative design multi-path residual block, constructed by parallel depth-wise separable convolution paths composed of convolution kernels of different sizes, is used to extract local minute and global large features, effectively enhancing the capacity of MPRANet. Meanwhile, the attention block is formed by cascading channel attention and spatial attention modules to re-scale features in the channel and spatial dimensions sequentially. A Mixture of Data Augmentation (MoDA) strategy for meliorating MPRANet performance without increasing computational burden is proposed. MoDA makes full use of multiple pixel-domain data augmentation methods to raise the generalization of MPRANet. Qualitative and quantitative experiments on three test datasets show that the proposed MPRANet has obvious advantages over state-of-the-art thermal and RGB image SR methods for the preservation of details such as edges and textures.