Picking up our RON95 ANPR prototype: a build diary

I wanted to start keeping a build diary for our RON95 ANPR project. There’s already code in the repo, but I haven’t really shared the journey behind it here.

So this first entry is a catch-up on the existing prototype and what I want to improve next.

The project explores automatic number plate recognition for petrol station workflows. The current image-processing script uses YOLO to locate a plate, crops that area, and passes it to PaddleOCR to read the characters. It then sends the plate number and confidence score to a Django backend.

Written down like that, the flow sounds fairly straightforward. But there are several places where things can go wrong before a useful record reaches the backend.

Finding the plate is one part. Reading its characters correctly is another. Then there’s the question of what to do when the result isn’t reliable enough.

That last part is something I want to work on.

Right now, the script skips sending a result when the OCR confidence falls below its configured threshold. There’s a check in place, but the next question is: what happens to that case afterwards?

I’d like to explore a review flow where staff can inspect the image, correct the reading if needed, or mark it as unreadable. That’s a planned improvement, and I still need to work through how it should behave.

I also want to keep the scope clear. Reading a number plate doesn’t establish someone’s subsidy eligibility, and this prototype still has a lot to prove before it could support a real station workflow.

Looking through the repo history, there’s also plenty of work that doesn’t make an exciting demo: dependencies, backend settings, admin pages, model configuration, and documentation. I want to include those parts in this diary too.

This is a team project, with xfnx and Sang Kancil involved alongside me. As we continue, I’ll use the GitHub updates to share what changed, what we actually tested, and what still needs attention.

For anyone who has worked with OCR: how did you handle uncertain readings without overwhelming people with manual reviews?