Guided Unfolded Networks for Fast and Accurate Multi-Camera Thermal Super-Resolution

Thermal cameras are widely used in robotics, surveillance, and autonomous sensing systems. However, High-Resolution (HR) thermal sensors remain expensive, which limits their use in many practical deployments. This motivates algorithmic approaches that enhance spatial detail from Low-Resolution (LR) thermal sensors. To address this challenge, this work presents a novel framework for multi-camera thermal Super-Resolution (SR) based on deep unfolding and knowledge distillation. The proposed models reconstruct HR thermal images from multiple LR cameras using our forward imaging formulation that reflects the thermal acquisition process, where each camera captures a translated view of the scene followed by spatial downsampling that degrades image quality. Guided by this formulation, we introduce the Gradient Updated Reconstruction Unit (GURU), an iterative teacher model that enforces measurement consistency while jointly learning inter-camera translations. The reconstruction process is then mapped into a learnable unfolded architecture called the Unfolded Network with Integrated Translations (UNIT), enabling faster inference through a fixed number of reconstruction stages. Finally, the GURU-guided Unfolded and Translationless Student (GUTS) is obtained through distillation, producing a lightweight model that removes the translation learning module and therefore achieves even faster inference than UNIT due to its reduced complexity while maintaining reconstruction fidelity. Experiments on our publicly available multi-camera thermal dataset show that the proposed models produce clear visual improvements over the LR inputs and compare favorably with state-of-the-art thermal SR approaches. At the same time, the distilled model achieves very fast inference time under 3 milliseconds, making the framework suitable for real-time deployment with low-cost thermal camera arrays.