)
Assessing car damage with the use of image recognition
A framework that recognises damage and offers assistance in the event of an accident.
Let’s begin with an example
Clara, our reference user, is involved in a car accident on her way to work. Nothing serious, apart from the bureaucratic issues that have to be dealt with. Calmly, Clara opens her car insurance app and photographs her car. The images are processed in real time by an engine that uses advanced AI technologies.
As a result, Clara receives details of the damaged parts, a cost estimate of the damage to her vehicle and an indication of the nearest affiliated workshops to contact. Is all this really possible?
,allowExpansion)
Image Recognition to classify car damage and estimate costs
Data Reply has developed a framework based on Deep Learning techniques (specifically Transfer Learning and Instance Segmentation), Data Mining and Natural Language Processing capable of classifying input data, such as the photos taken by appraisers and the repair data recorded by car repair shops, providing a quick response on:
the car and model recognised in the photos;
the location of the car components and the relevant extent of damage in each photo;
an estimate of repair costs based on the type of damage recognised.
The framework in more detail
In 90% of the verified cases, the model developed enables a correct recognition of the damage and an adequate estimate of the relevant repair costs.
Module 1: damage recognition
Damage recognition uses an set of convolutional neural networks trained on images containing different types of damage on cars of different brands and models. Pre-trained neural networks are used to leverage the potential of Transfer Learning, in addition to Instance Segmentation algorithms used to identify which car components have been affected. Once the training phase has been completed, the model facilitates the correct identification of which car components are damaged, as well as determining the severity of the damage present in the photo.
Module 2: cost estimation
To obtain the estimate of repair costs, the data provided by car repair shops is used, appropriately normalised and aggregated using data mining and NLP (natural language processing) algorithms. A statistical evaluation is subsequently made of the costs for repairing similar damage on the same car model. The estimates are then cross-referenced with the results of the damage recognition from photos, relying on a dedicated optimisation algorithm, in order to obtain an overall estimate of the accident repair costs.
A new customer experience for insurance companies and their customers
As illustrated in the initial example, the possibility of leveraging the framework developed by Data Reply in an app makes it possible to experiment with a new assistance method offered by a user’s insurance company, thanks to complete and real-time information.
Similarly, the insurance company can also benefit from using the app. How? By automating parts of the claims settlement process, it is possible, for example, to avoid the need for appraisers to be involved in some of the more common types of accidents. Moreover, communication between customers and car repair shops is facilitated. Finally, customers can be provided with tailor-made repayment solutions based on the various types of damage.
)
Data Reply is the Reply group company that specializes in big data, data science and artificial intelligence. We are building experience across four main business sectors: Sales and Marketing Intelligence, Big Data Engineering & Security Intelligence, Enterprise Intelligence, IoT & Industry 4.0 Intelligence. The more than 40 projects currently in production include the creation of Data Lakes and the use of Artificial Intelligence and Machine Learning algorithms. An innovative approach based on quantum computing supports the development of algorithms. We also offer training programmes on Data Science and Deep Learning.
,allowExpansion;Resize,width=660)
)
,allowExpansion;Resize,width=660)
,allowExpansion;Resize,width=660)
;Resize,width=660)