A very deep transfer learning model for vehicle damage detection and localization

Claims leakage is a major problem engendering tremendous losses for insurance companies. Those losses are due to the difference between the amount paid by insurance companies and the exact amount that should be spent, which cost millions of dollars yearly. Experts assert that these losses are caused by inefficient claims processing, frauds, and poor decision-making in the company. With the huge advances in Artificial Intelligence (AI), machine and deep learning algorithms, those technologies have started being used in insurance industry to solve such problems and cope with their negative consequences. In this paper, we propose automated and efficient deep learning-based architectures for vehicle damage detection and localization. The proposed solution combines deep learning, instance segmentation, and transfer learning techniques for features extraction and damage identification. Its objective is to automatically detect damages in vehicles, locate them, classify their severity levels, and visualize them by contouring their exact locations. Numerical results reveal that our transfer learning proposed solution, based on Inception-ResnetV2 pre-trained model followed by a fully connected neural network, achieves higher performances in features extraction and damage detection/localization than another pre-trained model, i.e., VGG16.