Geometry-referenced 3D mapping of thermographic anomalies on bronze objects using thermography-oriented super-resolution

Infrared thermography reveal subsurface anomalies in heritage objects, but mapping thermograms onto 3D surfaces remains challenging owing to limited thermal-image resolution and cross-modal differences between thermograms and geometric models. We integrated pulsed thermography, thermography-oriented super-resolution, laser scanning, and 2D-to-3D registration for geometry-referenced anomaly mapping. Fine-tuned with absolute-error loss, SwinIR yielded a peak signal-to-noise ratio of 46.00 dB and a structural similarity index of 0.9755 on a synthetically degraded validation data, slightly outperforming bicubic interpolation and producing fewer high-frequency artifacts than a generative adversarial network model. Normalized single-view reprojection errors were similar for raw and super-resolved inputs, showing that resolution alone did not reduce relative registration error; manual correspondence selection introduced uncertainty. Multi-view alignment to the laser-scanned model yielded a 2.215-pixels reprojection root-mean-square error in image space. On a bronze zun , the workflow mapped a thermographic anomaly associated with a visible surface crack, demonstrating the feasibility of geometry-referenced mapping and visualization.