Skip to main content




Artificial Intelligence
Machine Learning

Key Outcomes

  • Computer Vision
  • Deep Learning
  • Multimodal Problem
  • Post Processing
  • Autocorrection

Key Products & Services

  • Computer vision for object detection
  • Statistical modelling and K nearest neighbours
  • AWS (Sagemaker, EC2 with GPU, OpenSearch)

C Suite Level Results

Developed a solution able to accurately recover half of the images from being sent to manual processing, without increasing the risk of charging the customers incorrectly.

"We were able to half the number of images sent to manual review"

The Client

About The Client

Our client is one of the world’s largest transport companies, building and operating urban toll roads with 2 million daily trips on 20+ roads across 3 countries. We helped them develop their AI/ML capability with the aim to operationalise their solutions in near future.

The Problem

As a future driven organisation, our client made a strategic investment into the development of a technology solution to realise the business value of their data assets. They were seeking a partner who could help them build and operationalise a licence plate recognition software capable of reducing the manual handling overhead.

The client utilises the electronic and video tolling technologies to accurately identify and bill road users. In cases when a transponder device is not available and a licence plate read is not reliable, the road usage events are directed for manual identification. Typically, these images are more challenging to read.

Being able to reduce the volume of images sent for manual review without introducing errors is of considerable value to the client and places them well to respond to increased volumes of road trips.

Mantel Group were engaged to design and build a solution to recognise licence plates in challenging images that would also provide an improved confidence estimation.

The Solution

While the automatic licence plate recognition is a trivial task, the complexity of the project comes from the challenging images that are rejected and sent for manual review. These images have poor quality, defects or occluded characters making the correct recognition of licence plates challenging.

At the start of the project Mantel Group performed a thorough review of the performance and suitability of the commercial off-the-shelf solutions to solve the client’s problem. It was estimated that a custom built model would provide value beyond commercial solutions.

The team completed an in-depth analysis of the data to identify areas of further improvement. We leveraged the client’s domain knowledge and helped them realise the value of their prior work by training improved computer vision based licence plate detection and character recognition models.

The Mantel Group team further improved the recognition rate of the solution by developing a bespoke post processing algorithm to improve licence plate recognition post licence plate reading step. This algorithm enabled the autocorrection of incorrectly identified or partially read licence plates. This allowed the recovery of half of such images from being sent to manual review, while ensuring the risk of introducing errors is minimised.

Throughout the duration of the project, Mantel Group promoted and upheld highest industry standard ML engineering practices, ensuring the custom Licence Plate Recognition model would be operationalised successfully in a production environment.


Together with the client’s team, Mantel Group developed a solution able to accurately recover half of the images from being sent to manual processing, without increasing the risk of charging the customers incorrectly.