Car Damage Identification and Categorization Using Various Transfer Learning Models

Abstract

Cars play a critical role in today's modern world, and the ability to automatically classify car damages is of specific importance to the auto insurance industry. On a daily basis, car insurance companies cope with car inspection and testing. These inspections are labour-intensive, time-consuming, and sometimes inaccurate processes. Clients and insurance firms alike face costs and discomforts as a result of certain processes. Term strategy to assist, boost, or improve such strict monitoring processes may be feasible with current technology, even if entire modification is still a long way off. This research work analyzes the problem of automatic car damage detection and classification - this is an issue of importance to insurance companies in handling auto insurance claims quickly. In the classification of images, object recognition and image segmentation, developments in computer vision algorithms that employ deep learning have generated promising results. Convolutional Neural Networks (CNN) can be used for detection, analysis and estimation of various types of damage in different parts of the car. This research work has used the transfer learning-based models, Inception V3,Xception, VGG16, VGG19, ResNet50, and MobileNet in the Kera's library to train our model to predict the damage and to compare the efficacy of these models. The proposed dataset is trained with these pre-trained models in order to obtain the maximum accuracy and speed with negligible loss so that the model could be employed in real-life to predict the claim. Our analysis indicates that MobileNet is more accurate and the training speed is also less when compared to other models. An accuracy of 97.28% in predicting damage and classifying it into different types was achieved - this is substantially better than results achieved in the past in a similar test set.