Legentic is empowering insurance professionals with insights from open vehicle data, offering historic market life cycles and market values. Our technology also aids in locating stolen vehicles, making Legentic Platform a crucial tool in the battle against insurance fraud. Check out how we leverage AI and machine learning to make vehicle data access smarter and more efficient.
The single most identifiable feature of a vehicle in an image is the license plate. When our users seek information about a specific vehicle in our data, they can search by the plate number to see if and where we've identified that vehicle. This is our most noticeable application of artificial intelligence for our customers. While it may seem simple at first glance, we've put a lot of thought into how we do this in practice.
A million plates a day
Legentic scans license plates from about a million images daily, recording the discovered plates so that our customers can query them quickly and efficiently. Given the scale of our automatic license plate recognition (ALPR), we must do it cost-effectively.
Our advanced recognition system
Our in-house license plate recognition system is built on a combination of neural networks adapted to our specific use case from state-of-the-art open-source models. It consists of two main components: a license plate detection model and a character recognition model.
First, we scan images for license plates using a customized object detection model. These plates can be at various angles; they may be dirty, snowy, bent, or deformed, and the lighting conditions can vary greatly. Plates are not always even rectangular. For instance, in the Canadian Northwest Territories and Nunavut, some plates are shaped like a polar bear! Depending on the country, plate designs can differ significantly in color. This diversity means we must carefully train our object detector to discern what is and isn't a license plate.
Second, if plates are detected in the first step, we zoom in on each one and use a customized optical character recognition (OCR) model to read the plate number. Here, too, we face challenges: Fonts and colors vary worldwide, and in some countries, like the United Kingdom and Finland, the letter I is indistinguishable from the digit 1 and the letter O from the digit 0. In the United States, there are over 8000 plate designs, many with colorful picture backgrounds. Countries like Canada and Australia also feature a variety of plate designs.
Decoding the details
There can also be letters or digits on the plate that are not part of the plate number. For example, in many European countries, a plate often has the country code on a blue background with the EU star pattern or a national flag. In North America, plates commonly include the state or province name and a motto—such as "Live free or die" in New Hampshire. Plates may also have coats of arms, stickers, other markings, or a plate frame with the dealership's name. As part of our algorithm fine-tuning, we must teach it to ignore everything that is not part of the plate number while reading it.
Integrating AI into our data pipeline
We have integrated all this into our data pipeline, combined with heuristics, to decide which of the millions of images we process daily are worth scanning for plates. In the end, the plate numbers detected from images are stored in our search index, allowing our users to quickly find them via the Legentic Platform or API.
By leveraging advanced AI and machine learning techniques, Legentic provides a robust and efficient way for users to access vehicle data through license plate recognition. Our technology simplifies the search process, making it quick and reliable, even when dealing with a massive volume of images and diverse plate designs.
Want to know more about the Legentic Platform and how we can help you boost your investigations?
Reach out to Conor, for a quick meeting.