Vehicle Damage model classification for Zimbabwe Insurance Sector using MobileNetV2 and DenseNet121

Abstract

According to the United Nations Road Safety Performance Review-Zimbabwe report, every 15 minutes, five people die in road accidents within Zimbabwe, recording the highest number of accidents in the SADC region. The situation has brought more pressure and work in the insurance sector as they are expected to process all the claims accurately and timely. Deep learning entails automation, enhancement, analysis, and high accuracy in areas like speech recognition, object detection, and language translation. In this paper, two modern deep learning algorithms MobileNetV2 and DenseNetV121 were used to develop the vehicle damage classification models. The models were used to detect damaged main features of a car, which are: the door, bumper, windscreen, tail lamp, and headlamp. Mobile NetV 2’s53 layers and DenseNet121’s121 layers produced high accuracy rates for identifying damaged parts in vehicles. However, DenseNetV2 produced a higher accuracy of 84& than MobileNetV2, with an accuracy rate of 78%. The models also used low computational resources than the traditional algorithms making them applicable in different insurance companies as they can be easily embedded into client’s mobile phones.