Deep-Learning-based Dent Detection of Aircraft Surfaces using Synthetic Data

Abstract

Detecting and localizing dents on aircraft surfaces is crucial for maintaining their structural integrity. However, this task can be challenging for humans as dents are not very prominent to the naked eye and require the assistance of light reflections to reveal the damages across the surface. Latest state-of-the-art technologies such as lasers or cameras digitize this step, however the work load is shifted in identifying the dents to the virtual image. The integration of deep-learning methodologies can help automate dent detection. This study compares two object detection architectures: You Look Only Once (YOLOv11) and Real-Time DETection TRansformer (RT-DETR) for dent detection. A high quality dent dataset is prepared, consisting of real and synthetic images of common long- and mid-range aircraft fuselages, to train and test the models. The results indicate that YOLOv11 marginally outperforms RT-DETR in detecting dents with a mean accuracy precision (mAP50) score of 0.66 against the mAP50 value of 0.57 for RT-DETR.