Aircraft Dent Detection Utilizing Specular Reflections and Deep Learning

Abstract

The aviation industry requires frequent and thorough visual inspections to find and evaluate defects, which are expensive and time-consuming. Automating parts of the inspection process using robots and deep learning has the potential to improve speed and performance. Dents are the most difficult type of defect to detect, since they usually lack distinguishing colors, and can only be seen via shadows or reflections. To make dent detection easier and more reliable for deep learning models, a capturing setup utilizing specular reflections is tested. This necessitates the creation of a new dataset of aircraft surface images, where dents are made visible with the help of specular reflections. A new annotation method that makes use of an optical tracking system to automatically create annotations was developed to create the dataset. Two different models were trained on variations of the dataset and tested to determine their ability to detect dents in specular reflection images, and their viability for use in a robotic inspection scenario. This thesis shows that both RT-DETR and YOLOv12 have excellent dent detection performance on the new dataset, are fast and accurate when processing video, and can be suitably integrated into a robotic inspection setup.