[MRNAQA]

What

Make the rotated page C-contiguous before OCR in both table-orientation paths in pipeline_v2.py: whole-page tables and layout-detected tables. Covers 90, 180, and 270 degrees.

Why

np.rot90 returns a view. Passing that non-contiguous page to the OCR crop loop makes OpenCV convert the full page for each text box. Normalize once after choosing the rotation, before OCR, preserving orientation and pixels.

Related to #5178 and complements #5179, which covers the UVDoc producer.

Direction: correct the page-producing boundary before repeated crops; keep table orientation and OCR behavior.

Common practice: follows NumPy's rotated-view contract and contiguous conversion, using the same boundary correction as #5179.

Correction: refactor.

Cut: repeated downstream conversion of the rotated page; no code removed.

Add: one np.ascontiguousarray call in each rotated-page path, before OCR.

Validation

• Repository pre-commit hooks on the changed file: passed, including Black 24.4.2, Flake8 7.0.0, isort 5.12.0, license, import, and Python 3.8 checks.
• Executed each added normalization statement with synthetic arrays for k=1/2/3: both paths produce contiguous pages with identical page and perspective-crop pixels.
• Changed source compiles. No model inference or full OCR test was run for this PR.

Gain: the synthetic reproduction already posted on #5178 measured median crop time from 81.967041 ms to 0.203834 ms, with identical crop pixels (402.1x). It uses a seeded random 5600×4000×3 uint8 page, a 500×40 crop, five warmups, and 30 interleaved samples per layout. The one-time page conversion cost was 82.709770 ms, excluded from crop timing. Only k=1 was timed; this is not an end-to-end OCR speedup claim.

Benchmark environment: macOS 27.0 arm64, Python 3.12.13, NumPy 2.3.5, opencv-python-headless 4.10.0.84. The script requests one OpenCV thread; the build reports 18.

Test plan

Reproduction and pixel checks are described above. For model-level validation, run table recognition on a generated turned-table page at 90, 180, and 270 degrees, with and without layout detection, and compare crop pixels and OCR results against the base revision.

Compatibility

No public API, rotation, or model configuration changes. The zero-degree path is unchanged.

[CLAassistant]

All committers have signed the CLA.