Autoassess AI: Vision-Based Car Damage Assessment Using Deep Learning
page 5-31
Ma. Carmelle C. Pedrosa
Master in Applied Business Analytics
University of Asia and the Pacific
Brenda A. Quismorio, Ph.D.
Doctor of Philosophy in Business Administration
University of the Philippines, Diliman
Business Process and Analytics Lead, Corporate Planning and Review, University of Asia and the Pacific
Elmer Peramo
Lecturer, Master in Applied Business Analytics, University of Asia and the Pacific
Keywords: Automated damage assessment, Deep learning in automotive repair, Computer vision in car repair, YOLO, EfficientNet, MobileNet
ABSTRACT
This study addresses the need for an automated car damage assessment system to reduce processing time and increase throughout at the XYZ, repair center, a facility currently reliant on manual, time-intensive evaluations. To meet this need, the researcher developed a modular AI-based solution leveraging computer vision and deep learning, specifically targeting damage location, type, and severity classification. The solution utilized EfficientNetV2B0 and MobileNetV2 models for location and severity classification, and YOLOv8 (You Only Look Once) for damage type detection. These models were trained and evaluated on a curated dataset of car damage images, meeting technical success criteria for accuracy, F1 score, and mean Average Precision (mAP). EfficientNetV2B0 achieved a 92% F1 score for location classification, while YOLOv8s achieved a 74.3% mAP in damage type detection. YOLOv8s performed well with visually distinct damages but faced challenges with subtle damage types, such as cracks. This AI-driven solution shows potential for improving operational efficiency, aligning with XYZ’s goal of streamlining damage assessments, enhancing customer experience, and setting a new standard in automotive repair services.