# Page orientation check

Small CPU preprocessor used before PepperOCR-VL on photographed pages. It runs
the PP-LCNet_x1_0_doc_ori classifier (PaddleOCR) on every image, predicts
0/90/180/270 degrees, and writes an upright copy of each page to the output
directory. A page is rotated only when the predicted angle has confidence 0.85
or higher; otherwise it is copied unchanged.

```bash
uv run --project . --frozen python orient.py \
  --input-dir  /path/to/pages \
  --output-dir /path/to/upright_pages \
  --model-dir  ./model \
  --record-path /path/to/orientation.jsonl   # optional per-image log
```

`pyproject.toml` and `uv.lock` pin PaddlePaddle CPU 3.2.2, OpenCV, Pillow and
NumPy; `uv` creates the environment on first run. `model.sha256` lists the
checksums of the two classifier files.