[3A2] Evaluating image registration stitching techniques for partially overlapping ultrasonic corrosion maps
A Nichita and F Cegla
Imperial College London, UK
This work evaluates the performance of popular image registration (computer vision) techniques for stitching partially overlapping ultrasonic corrosion maps. The goal is to identify the conditions under which reliable stitching can be achieved and to quantify the success rate and resulting errors. We assess three feature detectors: KAZE, SIFT and SURF and two feature extraction formats: SURF and histogram of oriented gradients (HOG) on a dataset of 1200 simulated corrosion maps with 10 mm average thickness. Each map was split into two C-scan strips (100 × 200 pixels) to simulate phased array ultrasonic inspections, with overlaps ranging from 1 to 50 pixels and rotations up to 20°. The study also introduces inspection-related imperfections, including noise (0-2 mm) and encoder skips.
Results show that KAZE consistently outperforms SIFT and SURF, especially under noisy conditions (0.1-2 mm), achieving up to 98% stitching success when overlap is ≥10 pixels. While all detectors perform well (90%-95%) in error-free scenarios, performance degrades under higher noise and rotation. For extraction methods, SURF and HOG show similar overall performance. However, when depth variation is low (<1 mm), SURF excels with smaller overlaps (<20 pixels), while HOG is more robust when overlap exceeds 20 pixels. Among all inspection variables, large-angle rotations had the most detrimental impact on stitching accuracy, followed by high noise levels, with encoder skips having minimal effect.
Results show that KAZE consistently outperforms SIFT and SURF, especially under noisy conditions (0.1-2 mm), achieving up to 98% stitching success when overlap is ≥10 pixels. While all detectors perform well (90%-95%) in error-free scenarios, performance degrades under higher noise and rotation. For extraction methods, SURF and HOG show similar overall performance. However, when depth variation is low (<1 mm), SURF excels with smaller overlaps (<20 pixels), while HOG is more robust when overlap exceeds 20 pixels. Among all inspection variables, large-angle rotations had the most detrimental impact on stitching accuracy, followed by high noise levels, with encoder skips having minimal effect.
