Deep Learning-Based Car Damage Classification and Detection

Abstract

In this paper, we worked on the problem of vehicle damage classification/detection which can be used by insurance companies to automate the process of vehicle insurance claims in a quick fashion. The recent advances in computer vision largely due to the adoption of fast, scalable and end-to-end trainable convolutional neural networks make it technically feasible to recognize vehicle damages using deep convolutional networks. We manually collected and annotated images from various online sources containing different types of vehicle damages. Due to the relatively small size of our dataset, we used models pre-trained on a large and diverse dataset to avoid overfitting and learn more general features. Using CNN models pre-trained on ImageNet dataset and using several other techniques to improve the performance of the system, we were able to achieve top accuracy of 96.39%, significantly better than the current results in this work. Furthermore, to detect the region of damage, we used state-of-the-art YOLO object detector and achieving a maximum map score of 77.78% on the held-out test set, demonstrating that the model was able to successfully recognize different vehicle damages. In addition to this, we also propose a pipeline for a more robust identification of the damage in vehicles by combining the tasks of classification and detection. Overall, these results pave the way for further research in this problem domain, and we believe that collection of a more diverse dataset would be sufficient to implement an automated vehicle damage identification system in the near future.