First Stage Results
Here are some results from the first-stage detector on a plane wing. The first example demonstrates that our method can detect multiple dents along a laser line in a single pass. The second example highlights the detector’s ability to identify dents partially obscured by dirt—cases that are often missed when relying solely on visual appearance. However, the first-stage detector also has limitations. It cannot reliably distinguish true dents from other surface features such as rivets, screws, or gaps, which leads to a high number of false positives.


Sample results from the first-stage detector
To quantitatively assess the performance of our first-stage dent detector, we conducted an experiment by creating controlled dents on a metal plate using a hammer. The results showed a Precision of 65.7% and a Recall of 93.6%, indicating that our method effectively detects most dents but may also produce some false positives.

A sample image of mental plate and first-stage detection result
Second Stage Results
To evaluate the performance of the second stage, we conducted an experiment using patches extracted from the first stage. To assess generalizability, we designated images from one section as the test set and images from other sections as the training set, following an 80/20 train-test split. This setup yielded an accuracy of 91.8%. The resulting confusion matrix is presented in the image below.

Samples of Manually Labelled Patches

Confusion Matrix of the Second Stage
