cyberquantic logo header
EN-language img
FR-language img
breadcrumbs icon
Insurance

Manage claims

AI solution for car damage classification

AI solution for car damage classification

For:
Insurance companies, car owners/users
Goal:
Automate a Business Process
Problem addressed
To create an automated system for car damage classification using CNNs.
Experiment using transfer and ensemble learning to find which is better for
training a CNN for car damage classification.
Scope of use case
Car damage classification for common damage types such as bumper dent, door
dent, glass shatter, head lamp broken, tail lamp broken, scratch and smash.
Description
Today, in the car insurance industry, a lot of money is wasted
due to claims leakage. Claims leakage / underwriting leakage
is defined as the difference between the actual claim
payment made and the amount that is expected to have been
paid if all industry leading practices were applied. Visual
inspection and validation have been used to reduce such
effects. However, they introduce delays in claim processing.
There have been efforts by a few start-ups to mitigate delays
in claim-processing time. An automated system for car
insurance claim processing is the need of the hour. We
employ Convolutional Neural Network (CNN)-based
methods for classification of car damage types. Specifically,
45
we consider common damage types such as bumper dent,
door dent, glass shatter, head lamp broken, tail lamp broken,
scratch and smash. To the best of our knowledge, there is no
publicly available dataset for car damage classification.
Therefore, we created our own dataset by collecting images
from the web and manually annotating them. The
classification task is challenging due to factors such as large
inter-class similarity and barely visible damage. We
experimented with many techniques such as directly
training a CNN, pre-training a CNN using an auto-encoder
followed by fine-tuning, using transfer learning from large
CNNs trained on ImageNet and building an ensemble
classifier on top of the set of pre-trained classifiers. We
observed that transfer learning combined with ensemble
learning works the best. We also devise a method to localize
a particular damage type.
We achieved accuracy of 89,5 % with a combination of
transfer and ensemble learning. The same technique can be
used for localization of damage. Further, only car-specific
features may not be effective for damage classification. It
thus underlines the superiority of feature representation
learned from the large training sets.
We hosted the trained model on a cloud service that can be
plugged into applications using API and can be used for
automated first level assessment of damages in the car
insurance sector.
Interested in the same or similar project?
Submit a request and get a free project evaluation.