Assessing Car Damage with Convolutional Neural Networks
Harit Bandi, Suyash N. Joshi, Siddhant Bhagat, Amol Deshpande
Abstract
Manual estimation of damages in fields like construction, vehicular accidents has been the mainstay of the insurance business. However, such methods are replete with biases and inaccurate estimations. This paper deals with estimating car damage, primarily with auto insurers as our key potential customers. For this purpose, three distinct Transfer Learning approaches are used which detect the presence of damage, location, and severity of the damage. The basis for algorithms used lies in Convolutional Neural Networks, customized to optimize accuracy. Each approach is analyzed and varying degrees of accuracy were achieved across different models deployed ranging from 68% to 87%. Accuracy as high as 87.9% was obtained during the course of experiments. This research fine-tunes a number of existing approaches and opens doors for collaboration in image recognition, particularly for the car insurance domain.