Object detection is a critical component in safety-centric applications like autonomous navigation and remote sensing. However, system efficacy is significantly degraded by adverse weather, with performance dropping by up to 40% in hazy, foggy, or low-light conditions that reduce contrast in both visible (Red Green Blue) and thermal infrared imagery. This performance gap exists because detection models are typically trained on clear-weather data, limiting their real-world reliability. To address this, we introduce a comprehensive enhancement framework that leverages innovative deep learning architectures to improve object detection in challenging visibility. Our research presents a suite of novel neural networks, each targeting a distinct challenge: an end-to-end Red Green Blue dehazing network, a Mamba-based model for thermal dehazing, a physics-guided thermal enhancement framework, a temporal model for thermal video consistency, and a thermal-to-RGB colorization pipeline. Extensive validation confirms our methods' effectiveness, in some cases doubling the number of detectable objects in real-world scenarios. The impact of this work extends to wildlife conservation, surveillance, search and rescue, and medical imaging, providing a foundation for more robust and reliable perception systems.
Sargis Hovhannisyan (2025) studied this question.