OCR results too lossy(need help)

[mubashir-oss]

Bug description

Hello team, I’ve been experimenting with DocTR OCR, but I’m seeing inaccuracies: ‘AT&T Inc’ → ‘AT&TI Inc’. How can I improve the results?

Code snippet to reproduce the bug

### Error traceback

### Environment

need help

[felixdittrich92]

Hi @mubashir-oss 👋,

Sounds like there is an issue with the detection (does it slightly overlap ?).
Have you already tried different models we provide ?
Additional take a look here: https://mindee.github.io/doctr/using_doctr/using_models.html#advanced-options

lower value - up to 0.01 - more detections possible inaccurate
higher value - up to 0.99 - possible less detections more accurate

predictor.det_predictor.model.postprocessor.bin_thresh = 0.5

[mubashir-oss]

Hi @mubashir-oss 👋,

Sounds like there is an issue with the detection (does it slightly overlap ?). Have you already tried different models we provide ? Additional take a look here: https://mindee.github.io/doctr/using_doctr/using_models.html#advanced-options

lower value - up to 0.01 - more detections possible inaccurate higher value - up to 0.99 - possible less detections more accurate

predictor.det_predictor.model.postprocessor.bin_thresh = 0.5

so basically i am using parseq + dbresnet50 combination and some other incorrect words it gave was AL instead of AI