Automated image stitching of ultrasonic C-scan thickness data

This study investigates the performance of established image registration methods for stitching partially overlapping ultrasonic C-scan corrosion maps, with the aim of improving data continuity in phased array ultrasonic testing (PAUT). Three popular feature detectors, KAZE, SIFT, and SURF, are evaluated alongside two extraction methods, SURF descriptor format and Histogram of Oriented Gradients (HOG), on a dataset of 1,200 simulated corrosion maps with a 10 mm initial backwall thickness. Each corroded map is divided into two overlapping C-scan strips (100 × 200 pixels), simulating inspection scenarios with overlap ranging from 1 to 50 pixels and misalignments up to 20° rotation. The analysis incorporates realistic acquisition imperfections, including signal noise (0–2 mm) and encoder skips. Results show that KAZE outperforms SIFT and SURF, particularly under noisy conditions, achieving stitching success rates up to 98% when overlaps exceed 10 elements. In ideal conditions, all detectors maintain high performance (90%–95%), but noise and rotational misalignment significantly degrade the results. The two extraction methods are similar for large depth variations, but when depth variation is small ( ≤ 1 mm), SURF is more effective at small overlaps ( ≤ 20 elements), while HOG exhibits greater robustness at larger overlaps. Among all tested inspection errors, rotational misalignment had the greatest negative impact on stitching accuracy, followed by noise, with encoder skips showing minimal effect. These findings support the integration of advanced feature-based registration techniques into PAUT data processing workflows, enabling improved corrosion mapping and defect characterization across multiple scans.