Infrared imaging systems face significant challenges, including low contrast, limited detail resolution, and indistinct edges due to environmental conditions and sensor limitations. These issues, combined with high noise levels and various artifacts, substantially reduce the utility of thermal images across applications. Although conventional and deep learning-based enhancement methods show promise, they often struggle to handle diverse scenarios and consistently produce high-quality results. This paper presents a novel approach to thermal image enhancement that improves contrast and detail while maintaining practical applicability. Our key contributions include: (1) the development of a pioneering physics-guided thermal image decomposition network explicitly grounded in the modified Stefan-Boltzmann law, (2) implementing a specialized reflection-suppression subnetwork that targets reflection-induced false signatures and effectively mitigates the associated blurring and distortion artifacts common in thermal imaging systems; and (3) validating our approach through comprehensive computer simulations and ablation studies on benchmark datasets. Quantitative evaluations across multiple benchmark datasets (LTIR, autonomous vehicles, CVC-14, solar panels, and breast data sets) demonstrate that the proposed approach outperforms existing infrared enhancement methods. Compared with eight state-of-the-art techniques, our model achieves three times higher EME, highlighting its exceptional ability to enhance thermal contrast without compromising fine details. In addition, qualitative assessments reveal superior visual clarity and robust noise suppression, underscoring the many advantages of our approach for thermal imaging applications.
Hovhannisyan et al. (Thu,) studied this question.