Skip to main content

Improving Machine Learning to Read License Plates for Increased Toll Enforcement


Mayank Gupta, Conduent India Labs

The world of transportation and the solutions available to tackle transportation challenges have advanced rapidly over the past few decades. We are closer than ever to realizing transportation systems with self-driving cars, artificial intelligence (AI)-based traffic control and more automated tolling. One of the critical parts of AI-based traffic control, tolling and law enforcement is an automated method of recognizing vehicles. This is easily done through automated license plate recognition (ALPR), a technology that Conduent is working towards perfecting.

Traditionally, ALPR uses optical character recognition (OCR) to recognize individual characters extracted from the license plate using object detection and segmentation techniques. Even though modern OCRs have been perfected by the industry, most commercial ALPR systems do not give accuracies of over 90%. This is because the error rate of an OCR on a ten-digit license plate is compounded tenfold. This 10% inaccuracy is a huge inconvenience as it can lead to incorrect billing or imposition of fines.

Solving for the accuracy challenges of ALPR

License plate re-identification (not recognition) can solve this problem. Re-identification uses deep neural networks to extract “signatures” of each license plate image captured. These signatures are then matched against a database of signatures obtained from previously seen license plates. This database is effective, as more than 40% of the traffic is typically comprised of repeat customers.

This method of re-identification relies on the vehicle to have been seen before, so in other words it can be considered a “one-shot” re-identification where we get to see each license plate once before re-identification is done. The addition of this one image to the signature database usually requires OCR or human labelling at some level. Moreover, if this image turns out to be wrongly labeled, over-illuminated or poorly cropped, then the re-identification process fails every time that license plate is queried.

Evolving from one-shot to zero-shot re-identification

At Conduent Labs India, we are working on improving existing re-identification techniques by incorporating multi-modal matching into the idea. This means that we learn the ability to create a “text-signature” for the label of every license plate image in the training set. We then match this text-signature space to the image-signature space to be as close to each other as possible. Therefore it’s not necessary to see a license plate before it can be identified, hence the term “zero-shot re-identification.”  Instead, having just the number of the plate is all that’s needed to match an image in the database. Our method can reliably match a plate against any large database of regular tolling customers or stolen vehicles. Conduent is leveraging the power of cognitive advanced analytics and IoT to address the re-identification problem and as a result, help our clients in tolling and public safety make more accurate analysis and decisions.

Early data shows improved accuracy and speed

In tests, our zero-shot re-identification process achieved an accuracy of 99.6% for a database size of 40,000 plates. Even on a simulated database of 5 million plates, our re-Identification accuracy was 96%. Moreover, this method can extract a “signature” from a license plate image 200 times faster than using a deep neural network. Even matching of the signature against a database is faster because the zero-shot license plate re-identification method uses signatures of much smaller dimensions. Finally, our method can also be used as an extremely fast and low complexity ALPR system with an accuracy of 84%.

The zero-shot license plate re-identification method is highly accurate, fast, scalable and efficient. It does not rely upon having a manually labelled image database and can be easily transferred from one dataset to another, as long as the plates are of the same country. This means that it can push up ALPR accuracies at tolling booths. It also does not consume much space or computational power and is portable. This can help in places where network connectivity is an issue. The work is closely related to technology like Face Re-ID and can even work offline on police dashcams with a stored database of wanted plate numbers, for example. However, it should be noted that re-identification is not always an alternative to recognition. It is a supplemental tool to boost recognition accuracies.


Click here to learn more about artificial intelligence, computer vision, analytics, IoT, multi-modal matching research innovation and other technology-led digital interactions and how they can help you.