Thermal Homography in Photovoltaic Panels: Evaluating Deep Learning and Feature-Based Methods

Thermal imaging has become a vital tool for analyzing temperature variations in various fields, including medical diagnostics, industrial inspection, and environmental monitoring. However, the application of homography techniques in the thermal imaging domain, particularly for photovoltaic (PV) panels, remains underexplored. This paper presents a comprehensive evaluation of state-of-the-art deep homography models, specifically HomographyNet, in comparison with traditional feature-based methods such as ORB+RANSAC and SIFT+RANSAC. These methods are applied to close-range thermal images of PV panels, and their performance is assessed using the mean average corner error (MACE) metric. Extensive experiments analyze the accuracy and robustness of these techniques, as well as the influence of colormap representations on model performance. Key contributions include the evaluation of homography methods for thermal imaging, an in-depth analysis of colormap effects, and the introduction of a novel high-resolution thermal image dataset for PV panels. Results demonstrate that HomographyNet outperforms traditional methods, achieving a MACE of 13.65 with the Noon dataset compared to 24.78 with the Morning dataset. HomographyNets superior performance with magma colormap makes it particularly useful in PV applications such as identifying microcracks or hotspots in solar cells, where accurate thermal image alignment can enhance the visualization of subtle yet critical temperature anomalies.