The power of artificial intelligence in vehicle damage assessment: real-time detection of scratches and dents

Imagine that you’ve just bought a used car, exactly the model you’ve been looking for for months that meets all your expectations. After the purchase, you first need to take out insurance and most importantly, you’re looking to do it quickly and easily. Until now, most insurance companies offered the possibility of going to an official facility for an expert to then assess the condition of the vehicle. In recent years, we have seen that the major companies also allow photographs to be taken of the car’s body. These images are then sent for manual verification before the final policy is confirmed, a process that usually lasts several days. This reality is set to change in the coming months. MAPFRE has found the most agile and secure solution thanks to artificial intelligence (AI).

In the coming weeks, the Spanish insurance company is going to launch a pilot with clients for automatic damage verification via an algorithm, available 24 hours a day 365 days a year, applied to taking out used vehicle insurance. What does this breakthrough mean? It means that you can demonstrate the condition of your car’s body without having to travel anywhere, without needing expert or specialist intervention and you can activate your policy instantly. “Automatic damage assessment was the missing piece of the puzzle we needed to take a definitive step forward and exploit the true potential of digital photo verification, which started in 2019,” says Alberto Rodríguez, Technical Manager of the Private Automobile Area at MAPFRE ESPAÑA.

Step-by-step with your cellphone

User operation is comfortable and, above all, simple, providing immediacy and the possibility of having insurance with extended coverages at any time. When the client completes their online insurance purchase, they receive an SMS message with a link, which opens a web solution on their smartphone to guide them in taking pictures from different angles: right and left side, front and rear. The novelty introduced in this pilot is that, at that moment, a deep learning algorithm (a cutting-edge technique) analyzes the images in real-time and identifies the presence or absence of any damage. This assessment allows the customized policy to be adjusted to the client. If the client agrees with the assessment, they can then begin to enjoy the benefits of their insurance without having to wait or travel anywhere. And all this without human intervention.

Supervised training

How is it possible, for example, for artificial intelligence to distinguish a simple stain from a scratch? With training, a lot of training. This is where the collaboration between ControlExpert and MAPFRE was essential. The German technology partner provided a starting model that had already been pre-trained with millions of photos of different models of vehicles from different countries. The Spanish insurer then provided thousands more images (processed anonymously) of vehicles corresponding to real insurance policies from recent years, i.e. images that are representative of the type in this case study in Spain. “In order to train the predictive model, you need to teach it first. To do this, you feed it with a very high number of varied images, as well as the expected response for each of them, in other words, the reality of the damage. This way, after this learning process, you can give the model a photo that it has never seen and it will be able to determine whether or not the vehicle has been damaged. The results we have obtained are spectacular and enable selective automation with an accuracy of 95 percent,” explains Daniel Almodóvar, Innovation Delivery Manager at MAPFRE.

Transversal teams from the two companies participated in the development of this project, bringing specialists in data analysis, business and innovation together with engineers from CESVIMAP (the MAPFRE R&D center), experts in reports, etc. For Andreas Witte, the CTO and Technology Manager at ControlExpert, it has been a source of pride to “develop strong cooperation with MAPFRE and support it to be at the forefront of the insurance industry, providing it with the latest AI-driven products and services.” Witte confirms that ControlExpert’s products, developed in close cooperation between its research department and international universities and research institutes, are based on leading technologies—for example, Generative Adversarial Networks (GANs) and Instance Segmentation—that manage to deploy to the greatest extent accurate predictive models for their clients. The digital assessment of damage using AI can have a positive effect on insurance in a market where tens of thousands of policies are taken out every year. We’re talking about more than 1.8 million used car transactions in 2020, a figure that is expected to increase by 4 percent this year, according to data from the consultancy firm MSI for the car dealership representative Faconauto.

A very high level of accuracy

Analyzing this number of images to refine the model has taken a huge effort. For example, at first, when there was a bad quality photo or when it was poorly lit, the artificial intelligence detected dents or marks on the car body from the image of the parking sensors that many modern models have. To achieve a very high level of accuracy, the AI has needed to learn all those patterns. According to MAPFRE sources, two-thirds of the policies taken out are for undamaged cars and one-third have some kind of dent or mark. And out of these, half have low-severity damage, which is not so easy to notice.

Automatic image assessment with AI is part of MAPFRE’s strategic initiatives as an insurance company, as well as those of MAPFRE Open Innovation (MOI) as the company’s innovation division. With the collaboration of ControlExpert, an influential technological partner in the insurance market, MAPFRE has created a unique and personalized system that is a trailblazing solution in Spain for taking out used vehicle insurance and has enormous potential to be applied in other areas. The future of AI in the industry is promising.

Beyond automobile insurance, systems like this one can also be applied to taking out homeowners or commercial insurance. The key to these next-generation products is to streamline processes and provide the client with a personalized, simple and modern experience. Ultimately, the goal is to offer the client a better experience and more modern management, appropriate for the new times we live in.