A Novel Image Inpainting Approach for Geometric Calibration of Low-Cost Handheld Thermal Cameras

This paper presents an integrated calibration framework for consumer thermal cameras whose firmware-embedded display overlays cause failure of conventional corner detection algorithms. The approach combines Navier-Stokes image inpainting for automated overlay removal with UV-printed emissivity contrast targets fabricated on aluminium composite substrates. Under controlled masking of overlay-free imagery, restoration recovers 98.4% of checkerboard corners to within 0.048 px of their ground truth positions, while corners falling within the restored region carry a mean localisation error of 3.34 px and are identified for exclusion. Applied to a consumer handheld thermal camera, the framework raises calibration feasibility from complete failure without preprocessing to successful calibration in every configuration tested, achieving mean reprojection errors of 0.155-0.238 pixels, with held-out errors of 0.328-0.538 pixels on images not used in estimation, with both falling within the range reported in the literature for professional-grade systems. Bootstrap resampling establishes the repeatability of the estimated intrinsics, and hold-out validation confirms that the fitted model generalises to unseen views. Robustness is assessed across two substrate materials, three target scales and both indoor and outdoor environments, comprising 387 images. The complete preprocessing pipeline runs below 50 ms per image. This work demonstrates that computational preprocessing can overcome a hardware-imposed constraint previously assumed to preclude calibration of this class of device.