A multi-scale perception network for infrared thermal image super-resolution in welding

Infrared thermal image super-resolution (SR) techniques aim to reconstruct high-resolution images from low-resolution counterparts, which is crucial for various applications. However, existing SR methods typically utilize limited pixel information for feature extraction, leading to suboptimal reconstruction quality. To address this limitation, a multi-scale perception super-resolution (MPSR) method is proposed to leverage both spatial and channel-wise information for improved image enhancement. Capitalizing on the inherent color consistency of infrared thermal images, a novel multi-attention mechanism is proposed that integrates scale decomposition attention, cross-attention fusion, channel feature selection, and global perceptual fusion. This mechanism effectively exploits contextual information, enhancing the restoration of both high-frequency details and low-frequency structures, thereby achieving more accurate and refined image resolution. Furthermore, the IRAB-T dataset is introduced as the first to include infrared thermal images of typical industrial scenarios, such as cutting, milling, and welding, thereby facilitating the broader application of SR techniques in industrial settings. Extensive experiments on benchmark datasets demonstrate that MPSR outperforms existing state-of-the-art SR methods. Notably, MPSR achieves an average PSNR/SSIM gain exceeding 10.58/8.24% compared to the visible SR method SwinIR and over 10.11/11.29% compared to the infrared SR method IERN. Moreover, in real-world friction stir welding scenarios, MPSR enhances defect detection confidence by 15.15% for identifying the “Flash” defect, underscoring its practical utility in industrial image pre-processing stages. To facilitate further research and application, the code of the proposed MPSR will be made publicly available at https://github.com/TJU-IRA/MPSR.