Computer Vision for Damage Detection in Cars Images

This project presents the design and implementation of a car damage classification and detection pipeline aimed at automating the auto insurance claim process. Recent advances in computer vision, particularly the adoption of fast, scalable, and end-to-end trainable convolutional neural networks (CNNs), enable effective recognition of vehicle damage. To develop this system, we collected and manually annotated images from various online sources, leveraging a web crawler and utilizing a free open-source Kaggle dataset containing diverse vehicle damage types. To mitigate overfitting due to the relatively small dataset size, we employed pre-trained models on large and diverse datasets. By using pre-trained CNN models on the ImageNet dataset and applying preprocessing techniques, the system achieved an accuracy of $89 \%$, surpassing previous results on similar test sets. Additionally, a state-of-the-art Detectron2 object detector was employed to identify damaged regions, yielding a maximum score of $87 \%$ on the retained test set, demonstrating successful recognition of various vehicle damages. This work proposes a comprehensive pipeline that integrates classification and detection tasks, paving the way for further research in automated vehicle damage identification. The collection of additional data could enable the implementation of an effective automated system in the near future.