Detection and Localization of Car Damages Using Deep Learning

In the automotive industry, accurate and efficient vehicle damage detection is crucial for reducing the cost of claims and improving assessment procedures. This paper presents an automated method for vehicle damage detection using deep learning based You Only Look Once models, specifically YOLOv3, YOLOv5, and YOLOv8, in conjunction with two optimizers, AdamW and SGD. The COCOtoYOLO dataset contains pictures with six damage kinds. Gaussian Blur, Median Blur, Grayscale conversion, and Contrast Limited Adaptive Histogram Equalization (CLAHE) improved picture quality and feature visibility. The results reveal that YOLOv8, optimized with SGD over 200 epochs, achieved the highest mean Average Precision (mAP-50) of $72 \%$, surpassing other configurations. This study demonstrates the effectiveness of YOLOv8 and SGD in vehicle damage detection. The highest F1-score of ${9 3 \%}$ for glass shutter class claims the model’s robustness for different types of car damages.