Deep Learning-Based Car Damage Detection using YOLOv8

Abstract

This thesis presents a deep learning-based approach for automatic car damage detection, classification, and segmentation using computer vision techniques. In recent years, deep learning has emerged as a powerful solution for automating visual inspection tasks, espe- cially in the automotive sector where accurate damage assessment is essential for insur- ance, maintenance, and evaluation processes. The proposed work focuses on identifying and localizing different types of vehicle damages, including scratches, dents, broken glass, cracked parts, and damaged lights. To achieve this, state-of-the-art deep learning models such as YOLO, and DETR are implemented and analyzed. These models are widely used in object detection and image segmentation tasks due to their high performance and adaptability. The system is trained and evaluated using a custom dataset composed of Algerian car images, which reflects real-world conditions and adds a local context to the study. This dataset includes various damage annotations that allow for both de- tection and segmentation tasks. The main objective of this research is to compare the performance of different models using standard evaluation metrics such as mean Average Precision (mAP) and Intersection over Union (IoU). Furthermore, the study investigates how the use of a local dataset influences model accuracy and robustness. Ultimately, this work aims to contribute to the development of an efficient and practical system for automated vehicle damage assessment that can be applied in real-world scenarios.