How to improve OCR recognition accuracy

[tlbb764652810]

Question
I want to use EasyOCR, but the results are not very satisfactory, and I'm looking for a solution.

First of all, my test images are mainly in Japanese and English. I added DOCLING_SERVE_ARTIFACTS_PATH="", but the recognition quality is still not very good. However, the results are very good in the demo at https://www.jaided.ai/easyocr/. I'd like to know how to solve this issue.

Of course, if I can recognize the test images better, I don't necessarily have to use EasyOCR.

I would be extremely honored to receive your help.

test image:

parsed image:

easy ocr demo:

[dosubot]

There's a known issue where EasyOCR produces worse results when used inside Docling compared to using it standalone, particularly for non-Latin scripts like Japanese. Here are some things you can try:

1. Increase image resolution - Set images_scale=2.0 (default is 1.0), which significantly improves OCR accuracy:

pipeline_options = PdfPipelineOptions(
    do_ocr=True,
    ocr_options=EasyOcrOptions(lang=["ja", "en"]),
    images_scale=2.0,
)

2. Force full page OCR - Docling has a 75% bitmap coverage threshold that may prevent OCR from running on your image. Setting force_full_page_ocr=True ensures OCR processes the entire page:

pipeline_options = PdfPipelineOptions(
    do_ocr=True,
    force_full_page_ocr=True,
    ocr_options=EasyOcrOptions(lang=["ja", "en"]),
    images_scale=2.0,
)

When using force_full_page_ocr=True, export with traverse_pictures=True:

text = result.document.export_to_text(traverse_pictures=True)

3. Try Tesseract instead - Given the known EasyOCR quality issues within Docling, Tesseract may give better results for Japanese:

from docling.datamodel.pipeline_options import TesseractCliOcrOptions

ocr_options = TesseractCliOcrOptions(lang=["jpn", "eng"])

You'll need Tesseract installed with Japanese language packs (tesseract-ocr-jpn).

Could you share what settings you're currently using (API call or code) so we can pinpoint what might be missing?

To reply, just mention @dosu.

---

^{How did I do? Good | Irrelevant | Incorrect | Verbose | Hallucination | Report 🐛 | Other}

   

[tlbb764652810]

@dosu
I'm using docling-serve, but I noticed that in the API documentation, the default value of images_scale is 2. Also, since I'm starting docling-serve with Docker, I can only call the API.

[dosubot]

Good point — since images_scale is already 2.0 by default in the API, that's not the issue. For docling-serve via API, try adding force_full_page_ocr=true to your request. This forces OCR on the entire page rather than only on areas Docling identifies as needing it:

curl -X POST "http://localhost:5001/v1/convert/file" \
  -F "files=@your_image.png;type=image/png" \
  -F "to_formats=md" \
  -F "do_ocr=true" \
  -F "force_full_page_ocr=true" \
  -F "ocr_engine=easyocr" \
  -F "ocr_lang=[\"ja\",\"en\"]"

If results are still poor, you could try switching to Tesseract (use ocr_engine=tesseract with ocr_lang=["jpn","eng"]), but the default docling-serve Docker image may not include Tesseract — you'd need a custom image with it installed.

Could you share the exact API call you're making? That would help identify if there's a specific parameter missing.

To reply, just mention @dosu.

---

^{How did I do? Good | Irrelevant | Incorrect | Verbose | Hallucination | Report 🐛 | Other}