Hi everyone,
I’m working on a task that involves reading 9-character alphanumeric codes from small paper snippets like the one in the image below. These are similar to voucher codes or printed serials. Here’s an example image:
I have about 300 such images that I can use for fine-tuning. The goal is to either:
- Use a pre-trained model out-of-the-box, or
- Fine-tune a suitable OCR model to extract the 9-character string accurately.
So far, I’ve tried the following:
- TrOCR: Fine-tuned on my dataset but didn’t yield great results. Possibly due to suboptimal training settings.
- SmolDocling: Lightweight but not very accurate on my dataset.
- LLama3.2-vision: Works to some extent, but not reliable for precise character reading.
- YOLO (custom-trained): Trained an object detection model to identify individual characters and then concatenate the detections into a string. This actually gave the best results so far, but there are edge cases (e.g. poor detection of “I”) where it fails.
I suspect that a model more specialized in OCR string detection, especially for short codes, would work better than object detection or large vision-language models.
Any suggestions for models or approaches that would suit this task well? Bonus points if the model is relatively lightweight and easy to deploy.
