paddleocr

CLI for PDF text extraction using PP-OCRv5 models on ONNX Runtime. Runs 100% locally — no API keys, no network access, no PaddlePaddle dependency.

PP-OCRv5 is the #1 open-source OCR engine on OmniDocBench (CVPR 2025) with only 5M parameters — beating models 100x larger.

Why ONNX Runtime

This CLI uses the same PP-OCRv5 models as PaddleOCR but runs them on ONNX Runtime instead of PaddlePaddle. Same accuracy, dramatically lower resource usage:

Metric	PaddlePaddle	ONNX Runtime	Improvement
Peak RAM	16 GB	1.7 GB	9.4x less
Time per page	26s	3.4s	7.7x faster
RAM at load	1.2 GB	120 MB	10x less

Benchmark vs alternatives

Tested on 7 Colombian medical documents (SOAT, surgical records, anesthesia forms) against GPT-5.5 ground truth:

Engine	Word Recall	Word F1	RAM	Time/page
This CLI (ONNX)	64.9%	67.8%	1.7 GB	3.4s
PaddlePaddle (original)	69.9%	71.1%	16 GB	26s
Tesseract	39.3%	43.4%	~300 MB	~8s

On standard documents (clean printed forms): 96-99% word recall — identical to PaddlePaddle.

Requirements

• Python 3.10+
• ~2 GB RAM
• ONNX model files (see Model setup)

Install

pip install -e .

Model setup

Models are not included in the repo (92+ MB). The paddleocr setup command converts them automatically from PaddleOCR's cached PaddlePaddle models.

Prerequisites

1. paddle2onnx (conversion tool, Python 3.12 recommended — 3.13 not yet supported):

   pip install paddle2onnx>=2.0.1

2. PaddleX source models cached at ~/.paddlex/official_models/. If not present, run PaddleOCR once with PaddlePaddle to download them, or download manually from PaddlePaddle's model hub.

Convert models

paddleocr setup

This converts 4 models (detection, recognition, two orientation classifiers) to ONNX format. The command is idempotent — it skips models that already exist. Use --force to reconvert all models.

# Reconvert all models (e.g. after a corrupted conversion)
paddleocr setup --force

Usage

Extract text from PDF pages

# All pages to stdout
paddleocr ocr --path document.pdf

# Specific pages to stdout
paddleocr ocr --path document.pdf --pages 1,3,5

# Page range to files
paddleocr ocr --path document.pdf --pages 1-10 --output-dir ./output

# Parallel processing (2 workers)
paddleocr ocr --path document.pdf --workers 2 --output-dir ./output

PDF metadata

paddleocr info --path document.pdf

Export pages as images

paddleocr images --path document.pdf --pages 1,3 --output-dir ./pages
paddleocr images --path document.pdf --output-dir ./pages --dpi 150

Output

All text output is markdown. When using --output-dir, each page is saved as page_01.md, page_02.md, etc.

Models

4 PP-OCRv5 models converted to ONNX (installed via paddleocr setup):

Model	Size	Function
PP-OCRv5_server_det	84 MB	Text region detection
latin_PP-OCRv5_mobile_rec	7.7 MB	Latin text recognition (836 chars, 45 languages)
PP-LCNet_x1_0_doc_ori	6.5 MB	Document orientation (0/90/180/270)
PP-LCNet_x1_0_textline_ori	6.5 MB	Text line orientation (0/180)

Performance

Benchmarked on Apple M4 (10 cores, 24 GB RAM) with Colombian medical documents (SOAT), 300 DPI.

Metric	Value
Time per page	~3.4s
Throughput	~17 pages/min
RAM at load	~120 MB
RAM peak	~1.7 GB

Resource requirements

Resource	Minimum	Recommended
RAM	4 GB	8 GB
CPU	2 cores	4+ cores
Disk	~200 MB	~300 MB
GPU	Not required	Not required

OCR config

Validated on Spanish medical documents:

• Detection: PP-OCRv5 server, resize_long=960, stride 128, box_thresh=0.5, unclip_ratio=1.5
• Recognition: Latin PP-OCRv5 mobile, height 48px, 836-char dictionary
• Orientation: automatic document and text line rotation
• Input resize: 1/2 before OCR