Car Damage Detection with VGG-16

Abstract

In today's traffic accidents, assessing the damage and keeping damage records quickly is necessary. It is crucial to accelerate damage assessment studies to prevent traffic congestion caused by accidents and open the road to traffic quickly. Deep learning technologies provide various advantages in calculating the magnitude of the damage, displaying the damage situation, and making inferences about the material extent of the damage. In this study, a decision support system is aimed not only for insurance companies or official institutions to see the results but also for the end user and to reveal the damage class of the accidents. The software offered aims to provide an objective perspective not only in accident processes but also in quickly determining the financial value of the vehicle when buying and selling the vehicle. In this study, the training phase of our VGG16-based model, a CNN sub-model, was developed on the data set we obtained from the Kaggle platform (5757 images). With VGG16, our vehicle object detection rate is 98%, the accuracy rate of whether the vehicle is damaged is 90%, the results obtained in the training in which we detect the damaged area is 70%, and finally, the accuracy rate in determining the level of damage (low, medium and high) is 66% has been obtained.