Blind super-resolution reconstruction of infrared images based on dual-domain feature extraction

Abstract

Existing infrared image super-resolution (SR) reconstruction methods typically employ a single degradation model to characterize image degradation, which often fails to reconstruct high-quality clear images in real-world scenarios effectively. Compared to visible light images, infrared or infrared thermal imaging images have lower resolution, fewer textures and details, and are more susceptible to noise interference. To address the challenges associated with the reconstruction of infrared images under complex degradation conditions, this paper proposes a novel wavelet transform-based blind super-resolution network (WTB-Net). WTB-Net can effectively realize multi-type degenerate kernel estimation, alternating optimization, and subsequent super-resolution image reconstruction under a unified framework. Firstly, to better extract the feature and detailed features of infrared images, we designed a bi-domain feature extraction module (BDFE), which is composed of multi-scale frequency domain feature extraction blocks (FDB) and spatial domain feature extraction blocks (SDB). In addition, we propose an edge attention mechanism to improve the reconstruction of edges and details. The above features will be input into the restorer and estimator for multiple super-resolution reconstructions, and the reconstruction results will be applied to the estimation and updating of a multi-class degraded fuzzy kernel. Finally, with the input of the low-resolution (LR) image, the high-resolution (HR) image can be generated based on the degraded fuzzy kernel and reconstruction network to complete the blind super-resolution reconstruction process. The proposed method is systematically compared with both classical and common super-resolution reconstruction methods using the Airai Electro-Optics Database and the LLVIP Database. The qualitative and quantitative results consistently validate the effectiveness of WTB-Net.