ppocr-lite 0.2.0

Lightweight PP-OCR runtime – ONNX only, no OpenCV, no heavy frameworks

ppocr-lite

A lightweight PaddlePaddle-OCR runtime for images like screenshots.

Dependency	Role
numpy	All numerical computation
Pillow	Image I/O and resize
onnxruntime	Model inference
scipy (optional)	Faster connected-component labelling

No OpenCV, deep-learning framework or utility libraries.

---

Install

pip install ppocr-lite        # CPU
pip install ppocr-lite[gpu]   # GPU (uses onnxruntime-gpu)
pip install ppocr-lite[fast]  # + scipy for faster CC labelling

Models (PP-OCRv5 mobile det/rec + v2 direction cls) can be auto-downloaded to ~/.cache/ppocr_lite/ on first use, or manually downloaded and configured.

Automatically downloaded models come from RapidOCR and are downloaded from huggingface (see here for details).

To manually download models see their huggingface - you'll need one det.onnx (for text detection), one rec.onnx (for text recognition) and the corresponding dict.txt (the model-output-to-character mapping). The mobile (= smaller) models as shipped by OnnxOCR also work quite well.

---

Quick start

from ppocr_lite import PPOCRLite

ocr_engine = PPOCRLite()

for result in ocr_engine.run("screenshot.png"):
    print(f"{result.score:.2f}  {result.text}")
    # result.box is a np.ndarray (4, 2) - top-left, top-right, bottom-right, bottom-left

Use your own models

from ppocr_lite import PPOCRLite, ModelConfig
from pathlib import Path

ocr_engine = PPOCRLite(
    ModelConfig(
        det_model=Path("models/PP-OCRv5/det.onnx"),
        rec_model=Path("models/PP-OCRv5/rec.onnx"),
        dict_path=Path("models/PP-OCRv5/dict.txt"),
        cls_model=False,   # skip direction classifier
    )
)

GPU inference

ocr_engine = PPOCRLite(providers=["CUDAExecutionProvider", "CPUExecutionProvider"])

---

Design notes

This project is very similar to the excellent RapidOCR, but more lightweight. Notably, it does not depend on OpenCV (which weighs around 200MB) and uses numpy-based alternatives instead. This does not hurt performance much, at least in my humble tests.

Please be aware that many of those numpy-based alternatives are only really feasible because this project assumes non-distorted input images (screenshots, clean document scans, …). I have not tested it, but I'd assume it doesn't work nearly as well on inputs like perspective-distorted real-world photographs.

What's different here?

• Detection post-processing – contour finding is replaced with scipy ndimage.label (or a numpy fallback). The minimum-area rectangle is simplified under the assumption of non-perspective distorted input. Polygon offset ("unclip") is done analytically using the area/perimeter ratio and a per-vertex outward push — accurate enough for near-rectangular screenshot text.

• Resize – PIL BILINEAR instead of cv2.resize. The two are numerically equivalent for the precision required by OCR.

• Crop – axis-aligned bounding-rect crop instead of a perspective warp. Screenshot text is always axis-aligned, making this lossless.

• No config YAML, no omegaconf – plain Python dataclasses.

Limitations vs. full PaddleOCR

• No perspective correction
• Direction classifier is only a 0°/180° binary; no 90°/270° support.

---

License

This project is GPL-3.0-or-later licensed. Note that the licenses of models (self-brought or auto-downloaded) will likely differ; refer to their creators for more information.