Deep Learning Approach for Damage Detection on Composite Structures

Abstract This research introduces a new method for inspecting composite panels to detect Barely Visible Impact Damage (BVID) and Visible Impact Damage (VID) using computer vision algorithms. The ultimate objective of this paper is to create a new approach for visual inspections of aircraft composite parts, leveraging digital optics and deep learning techniques to identify and categorize damage, using computer vision algorithms as tools to automate surface inspection and damage detection processes. In particular, in this work, two state-of-the-art object detection architectures are compared: You Only Look Once (YOLO) and Real-Time DEtection Transformer (RT-DETR). The study provides an overview of the challenges, methods, and recent advancements in deep learning algorithms used for general damage detection. Data for training convolutional neural networks (CNNs) were gathered using a high-quality acquisition system. The database was built by collecting images of damage from various composite panels that were previously impacted by different objects. Each defect was photographed under varying lighting conditions (bright and dark), resolutions, and lighting angles to accurately replicate the environmental conditions of an aircraft maintenance hangar.