Car damage assessment recommendation system using neural networks

Abstract

Traditional Car dents severity prediction involves skillful experts in fields that tends to be tedious and time-consuming. Considering the copious amounts of accidents, a robust, automated, cost-effective approach for detecting severity in car dents deems necessary. Our main goal is to utilize modern procedures for processing accident severity to overcome the problems associated with human visual error-prone problems. Mostly, vehicle accidents have specific fixed set of parameters that play a crucial role in determining the severity of the accident. In our system, we incorporate an approach named transfer learning for the image feature extraction process via pre-trained models for predicting severity of the dent in the vehicle. Our system initially extracts car image samples which are pre-processed for background elimination. These images are later passed on to our pre-trained neural network models for fine-tuning and feature extraction. Finally, we generate, and test various categories of dents based on their severity with existing pre-classified samples to enable high similarity score prediction. Several experimental analysis and exhaustive testing, our proposed system was able to detect vehicle dent severity with an overall accuracy of 96.34%, which is ably better than other existing computer-based classification methods.