Meter OCR

Hello,

i want to train a model with over 90 accuracy to identfy meter readings. I found some really good datasets. Now my problems:

  • no decimal place in the datasets → i need decimal places, i would also train them with own images
  • which model is the best ? accuracy is more important then speed.

Thank you.

Hey @nycrom

It sounds like you might be trying to train an object detection model or OCR at the moment. Object detection models infer bounding boxes around objects of interest and might not be suitable for your use case. OCR reads text off images and can’t identify the pointer in a meter.

If you want to derive precise angles (which you can use to convert to meter readings), I recommend using a keypoint detection model.

A contributor on our forum recently wrote a guide on how to estimate bottle orientation using keypoint detection. I think that’s a good place to start, just substitute bottles for the meter pointer/hand.

You can also search for keypoint detection specific models using Universe:

Hey Leo,

thanks for the reply.

No, these are only electricity meters that have either analogue digits or digital digits.

Hey @nycrom

Oh, thanks for clarifying! In that case, I would recommend you use a object detection model to locate the meter itself, then use an OCR model (we’ve covered several here)

Hope this helps

Okay so i only should train a model on your platform to identify the meter and then use e.g. Claude? The problem with GenAi is, we have no confidence value. I need the confidence.

Hey @nycrom

We still need to update the post, but PaliGemma, a new multimodal model, does OCR the best and it can run locally on your device using Inference.

PaliGemma doesn’t provide confidence values either, but its worth noting that OCR-specific models and packages usually don’t provide confidence either.

If you really wanted to use object detection, you could do something like this: Digits Object Detection Dataset and Pre-Trained Model by Phils Workspace

Could you share why you need the confidence values for the OCR prediction?

1 Like

Thank you.

We have an automated process and if the confidence is too low, an agent looks at the case. If its good enough, the values will be processed in the ERP system. So confidence is crucial and also a good quality.

Hello!
@nycrom please look at this url Welcome to AI-meter-recognition-PHP Discussions! · logicland-public-repositories/AI-meter-recognition-PHP · Discussion #1 · GitHub . Is this what you want to achieve?
Here is a sample of recognition: