README.md

metadata

license: mit
language:
  - en
  - tr

PaddleOCR Mobile Quantized Models (ONNX)

Overview

This repo hosts four ONNX models converted from PaddleOCR mobile checkpoints

File	Task	Language scope	Input shape
Multilingual_PP-OCRv3_det_infer.onnx	Text-detection	80+ scripts	NCHW • 1×3×H×W
PP-OCRv3_mobile_det_infer.onnx	Text-detection	Latin only	1×3×H×W
ch_ppocr_mobile_v2.0_cls_infer.onnx	Angle classifier	Chinese/Latin	1×3×H×W
latin_PP-OCRv3_mobile_rec_infer.onnx	Text-recognition	Latin	1×3×H×W

All models were:

• exported with paddle2onnx 1.2.3 (opset 11)
• simplified via onnx-simplifier 0.4+

Quick Start

import onnxruntime as ort, numpy as np
img = np.random.rand(1, 3, 224, 224).astype("float32")

det   = ort.InferenceSession("Multilingual_PP-OCRv3_det_infer.onnx")
cls   = ort.InferenceSession("ch_ppocr_mobile_v2.0_cls_infer.onnx")
rec   = ort.InferenceSession("latin_PP-OCRv3_mobile_rec_infer.onnx")

det_out = det.run(None, {det.get_inputs()[0].name: img})[0]
# add your post-processing / cropping / decoding here …