Real-World Thermal Image Super-Resolution

Thermal cameras are used in various domains where the vision of RGB cameras is limited. Thermographic imaging enables the visualizations of objects beyond the visible range, which enables its use in many applications like autonomous cars, nightly footage, military, or surveillance. However, the high cost of manufacturing this type of camera limits the spatial resolution that it can provide. Real-World Super-Resolution (RWSR) is a topic that can be used to solve this problem by using image processing techniques that enhance the quality of a real-world image by reconstructing lost high-frequency information. This work adapts an existing RWSR framework that is designed to super-resolve real-world RGB images. This framework estimates the degradation parameters needed to generate realistic Low-resolution (LR) and High-resolution (HR) image pairs, then the SR model learns the mapping between the LR and HR domains using the constructed image pairs and applies this mapping to new LR thermal images. The experiments results show a clear improvement in the perceptual quality in terms of clarity and sharpness, which surpasses the performance of the current SotA method for thermal image SR.