Mineru 3.4.4 Regression?

[auzty]

I would to ask, i have a scanned PDF

this is the output with mineru 3.1.11 and it's very great :

but when i trying to do with mineru 3.4.4 even using cli
mineru -p input.pdf -o output --effort high -m ocr -b hybrid-engine
i get

the differences are, the underline is missing when using 3.4.4
*actually i'm okay with 3.1.11 but just curious why this is happened? or i need some adjustment at 3.4.4?
I run mineru from docker, and following the base images runpod/base:0.6.3-cuda12.1.0 and install the mineru inside it

thank you

[dosubot[bot]]

Hi @auzty! I'm Dosu and I'm helping the OpenDataLab team.

The quality difference you're seeing is most likely due to the default effort level changing between these versions. MinerU v3.3 introduced medium and high effort levels, with medium as the new default [1]. In v3.1.11 there were no effort levels — the engine always ran at full quality. So even though you're passing --effort high in your CLI command, there are a couple of other changes that may also contribute:

1. OCR detector thresholds were tightened in v3.1.15 — det_db_box_thresh was raised from 0.3 → 0.5 and det_db_unclip_ratio reduced from 1.8 → 1.5 [2]. This can reduce text-region recall on scanned PDFs with faint or tightly spaced text, which could affect how well underlined text and formulas are detected.

2. The layout detection model changed from doclayoutyolo to PP-DocLayoutV2 between these versions [3], which introduced inline-formula detection/masking and new block types. While generally an improvement, it can change how text regions and formula regions are carved out on scanned pages.

3. Dense layout handling has documented issues under medium effort in v3.4.x, with missing text blocks reported and confirmed by maintainers — switching to high resolves it [4] [5].

Since you're already using --effort high, the regression you're seeing is likely coming from the OCR threshold changes and/or the layout model swap. You could try:

• Using pipeline backend instead of hybrid-engine (-b pipeline-engine) to see if the pipeline-only path handles your scanned PDF better for formula/text extraction, since it uses a purely recognition-based MFR model without generative components [6].
• Comparing the middle.json output between both versions to see whether the issue is in layout detection (text blocks not detected) or OCR/formula recognition (blocks detected but content is wrong).

Would you like me to dig deeper into the specific layout or OCR changes that might affect underline detection in scanned documents?

To reply, just mention @dosu.

---

Docs are dead. Just use Dosu.