openocr-skill

Extract text from images, documents and scanned PDFs using OpenOCR - a lightweight and efficient OCR system with document parsing model requiring only 0.1B parameters, capable of running recognition on personal PCs. Supports text detection, recognition, universal VLM recognition, and document parsing with layout analysis

Security Scan

The skill's requests and instructions are consistent with an OCR/document-parsing helper — it expects to run local OCR code and download models, but it provides no installer or explicit network/model sources so verify environment and model origins before use.

Purpose & Capability

The name/description (OpenOCR OCR, document parsing, VLM recognition) matches the instructions: code examples show initializing OpenOCR for detection/recognition/doc parsing and working with image/pdf paths. There are no unrelated credentials, binaries, or config paths requested.

Instruction Scope

SKILL.md contains only OCR-related instructions and Python examples. It repeatedly relies on passing image/PDF file paths and on auto-downloading models; it does not instruct accessing unrelated system files or secrets. However, the instructions assume ability to run Python code, read/write files, and fetch model binaries from the network.

Install Mechanism

This is an instruction-only skill with no install spec. The skill examples expect a Python package ('openocr') and optional ONNX/Torch backends and include 'auto_download' of models. Because no install/source-of-models is declared, model and dependency downloads would be performed dynamically by the runtime — verify where those downloads come from and that you trust the source.

Credentials

The skill requests no environment variables, credentials, or config paths. It only needs file access to user-provided images/PDFs and network access to fetch models (if auto_download is used), which is proportional to OCR functionality.

Persistence & Privilege

The skill does not ask to be always-enabled, does not declare persistence or modify other skills, and requires no long-lived credentials. It does instruct writing/reading model and cache files locally, which is typical for model-based tools.

Assessment

This skill appears to do what it claims (local OCR and document parsing). Before installing/using it: (1) confirm you have a Python environment and the required ML runtimes (ONNX/Torch) or be prepared to install them; (2) expect the skill to auto-download model weights — check the OpenOCR project's GitHub and verify the model download URLs and integrity or supply model files yourself to avoid unexpected network fetches; (3) be aware the agent will need permission to read the image/PDF files you give it and to write cache/model files to disk; and (4) if you need higher assurance, ask the publisher for an explicit install spec and the exact model-hosting URLs (or use only local model paths).

Overview

This skill enables intelligent text extraction, document parsing, and universal recognition using OpenOCR - an accurate and efficient general OCR system. It provides a unified interface for text detection, text recognition, end-to-end OCR, VLM-based universal recognition (text/formulas/tables), and document parsing with layout analysis. Supports Chinese, English, and more.

How to Use

1. Provide the image, scanned document, or PDF
2. Optionally specify the task type (det/rec/ocr/unirec/doc)
3. I'll extract text, formulas, tables, or full document structure

Example prompts:

• "Extract all text from this image"
• "Detect text regions in this photo"
• "Recognize the formula in this screenshot"
• "Parse this PDF document with layout analysis"
• "Convert this scanned page to Markdown"

Domain Knowledge

OpenOCR Fundamentals

from openocr import OpenOCR

# Initialize with a specific task
engine = OpenOCR(task='ocr')

# Run OCR on an image (callable interface)
results, time_dicts = engine(image_path='image.jpg')

# Results contain detected boxes with recognized text
for result in results:
    for line in result:
        box = line[0]       # Bounding box coordinates
        text = line[1][0]   # Recognized text
        conf = line[1][1]   # Confidence score
        print(f"{text} ({conf:.2f})")

Supported Tasks

# Available task types
tasks = {
    'det':    'Text Detection - detect text regions with bounding boxes',
    'rec':    'Text Recognition - recognize text from cropped images',
    'ocr':    'End-to-End OCR - detection + recognition pipeline',
    'unirec': 'Universal Recognition - VLM-based text/formula/table recognition (0.1B params)',
    'doc':    'Document Parsing - layout analysis + universal recognition (0.1B params)',
}

# Task selection via parameter
det_engine = OpenOCR(task='det')
rec_engine = OpenOCR(task='rec')
ocr_engine = OpenOCR(task='ocr')
unirec_engine = OpenOCR(task='unirec')
doc_engine = OpenOCR(task='doc')

Configuration Options

from openocr import OpenOCR

# === Text Detection ===
detector = OpenOCR(
    task='det',
    backend='onnx',                          # 'onnx' (default) or 'torch'
    onnx_det_model_path=None,                # Custom detection model (auto-downloads if None)
    use_gpu='auto',                          # 'auto', 'true', or 'false'
)

# === Text Recognition ===
recognizer = OpenOCR(
    task='rec',
    mode='mobile',                           # 'mobile' (fast) or 'server' (accurate)
    backend='onnx',                          # 'onnx' (default) or 'torch'
    onnx_rec_model_path=None,                # Custom recognition model
    use_gpu='auto',
)

# === End-to-End OCR ===
ocr = OpenOCR(
    task='ocr',
    mode='mobile',                           # 'mobile' or 'server'
    backend='onnx',                          # 'onnx' or 'torch'
    onnx_det_model_path=None,                # Custom detection model
    onnx_rec_model_path=None,                # Custom recognition model
    drop_score=0.5,                          # Confidence threshold for filtering
    det_box_type='quad',                     # 'quad' or 'poly' (for curved text)
    use_gpu='auto',
)

# === Universal Recognition (UniRec) ===
unirec = OpenOCR(
    task='unirec',
    unirec_encoder_path=None,                # Custom encoder ONNX model
    unirec_decoder_path=None,                # Custom decoder ONNX model
    tokenizer_mapping_path=None,             # Custom tokenizer mapping JSON
    max_length=2048,                         # Max generation length
    auto_download=True,                      # Auto-download missing models
    use_gpu='auto',
)

# === Document Parsing (OpenDoc) ===
doc = OpenOCR(
    task='doc',
    layout_model_path=None,                  # Custom layout detection model (PP-DocLayoutV2)
    unirec_encoder_path=None,                # Custom UniRec encoder
    unirec_decoder_path=None,                # Custom UniRec decoder
    tokenizer_mapping_path=None,             # Custom tokenizer mapping
    layout_threshold=0.5,                    # Layout detection threshold
    use_layout_detection=True,               # Enable layout analysis
    max_parallel_blocks=4,                   # Max parallel VLM blocks
    auto_download=True,                      # Auto-download missing models
    use_gpu='auto',
)

Task-Specific Usage

Text Detection

from openocr import OpenOCR

detector = OpenOCR(task='det', backend='onnx')

# Detect text regions
results = detector(image_path='image.jpg')

boxes = results[0]['boxes']      # np.ndarray of bounding boxes
elapse = results[0]['elapse']    # Processing time in seconds

print(f"Found {len(boxes)} text regions in {elapse:.3f}s")
for box in boxes:
    print(f"  Box: {box.tolist()}")

Text Recognition

from openocr import OpenOCR

# Mobile mode (fast, ONNX)
recognizer = OpenOCR(task='rec', mode='mobile', backend='onnx')

# Server mode (accurate, requires torch)
# recognizer = OpenOCR(task='rec', mode='server', backend='torch')

results = recognizer(image_path='word.jpg', batch_num=1)

text = results[0]['text']        # Recognized text string
score = results[0]['score']      # Confidence score
elapse = results[0]['elapse']    # Processing time

print(f"Text: {text}, Score: {score:.3f}, Time: {elapse:.3f}s")

End-to-End OCR

from openocr import OpenOCR

ocr = OpenOCR(task='ocr', mode='mobile', backend='onnx')

# Run OCR with visualization
results, time_dicts = ocr(
    image_path='image.jpg',
    save_dir='./output',
    is_visualize=True,
    rec_batch_num=6,
)

# Process results
for result in results:
    for line in result:
        box, (text, confidence) = line[0], line[1]
        print(f"{text} ({confidence:.2f})")

Universal Recognition (UniRec)

from openocr import OpenOCR

unirec = OpenOCR(task='unirec')

# Image input
result_text, generated_ids = unirec(image_path='formula.jpg', max_length=2048)
print(f"Result: {result_text}")

# PDF input (returns list of tuples, one per page)
results = unirec(image_path='document.pdf', max_length=2048)
for page_text, page_ids in results:
    print(f"Page: {page_text[:100]}...")

Document Parsing (OpenDoc)

from openocr import OpenOCR

doc = OpenOCR(task='doc', use_layout_detection=True)

# Parse a document image
result = doc(image_path='document.jpg')

# Save outputs in multiple formats
doc.save_to_markdown(result, './output')
doc.save_to_json(result, './output')
doc.save_visualization(result, './output')

# Parse a PDF (returns list of dicts, one per page)
results = doc(image_path='document.pdf')
for page_result in results:
    doc.save_to_markdown(page_result, './output')

Command-Line Interface

# Text Detection
openocr --task det --input_path image.jpg --is_vis

# Text Recognition
openocr --task rec --input_path word.jpg --mode server --backend torch

# End-to-End OCR
openocr --task ocr --input_path image.jpg --is_vis --output_path ./results

# Universal Recognition
openocr --task unirec --input_path formula.jpg --max_length 2048

# Document Parsing
openocr --task doc --input_path document.pdf \
    --use_layout_detection --save_vis --save_json --save_markdown

# Launch Gradio Demos
openocr --task launch_openocr_demo --share --server_port 7860
openocr --task launch_unirec_demo --share --server_port 7861
openocr --task launch_opendoc_demo --share --server_port 7862

Processing Different Sources

Image Files

from openocr import OpenOCR

ocr = OpenOCR(task='ocr')

# Single image
results, _ = ocr(image_path='image.jpg')

# Directory of images
results, _ = ocr(image_path='./images/', save_dir='./output', is_visualize=True)

PDF Files

from openocr import OpenOCR

# UniRec handles PDFs natively
unirec = OpenOCR(task='unirec')
results = unirec(image_path='document.pdf', max_length=2048)

# OpenDoc handles PDFs natively with layout analysis
doc = OpenOCR(task='doc', use_layout_detection=True)
results = doc(image_path='document.pdf')

# Save each page
for page_result in results:
    doc.save_to_markdown(page_result, './output')
    doc.save_to_json(page_result, './output')

Numpy Array Input

import cv2
from openocr import OpenOCR

ocr = OpenOCR(task='ocr')

# Read image as numpy array
img = cv2.imread('image.jpg')

# Pass numpy array directly
results, _ = ocr(img_numpy=img)

Result Formats

# Detection result format
det_result = [{'boxes': np.ndarray, 'elapse': float}]

# Recognition result format
rec_result = [{'text': str, 'score': float, 'elapse': float}]

# OCR result format (detection + recognition)
ocr_result = (results_list, time_dicts)
# results_list: [[[box, (text, confidence)], ...], ...]

# UniRec result format
# Image: (text: str, generated_ids: list)
# PDF:   [(text: str, generated_ids: list), ...]  # one per page

# Doc result format
# Image: dict with layout blocks and recognized content
# PDF:   [dict, ...]  # one per page

Best Practices

1. Choose the Right Task: Use ocr for general text, unirec for formulas/tables, doc for full documents
2. Use Mobile Mode for Speed: mode='mobile' is much faster; use mode='server' only when accuracy is critical
3. Use ONNX Backend: Default ONNX backend works on CPU without extra dependencies
4. Set Appropriate Thresholds: Adjust drop_score (OCR) and layout_threshold (Doc) for your use case
5. Enable Layout Detection: For documents with mixed content (text + formulas + tables), always enable use_layout_detection
6. Batch Processing: Use rec_batch_num to control recognition batch size for throughput optimization
7. GPU Acceleration: Install onnxruntime-gpu or PyTorch with CUDA for significant speedup

Common Patterns

Full Document Processing Pipeline

from openocr import OpenOCR
import os

def process_documents(input_dir, output_dir):
    """Process all documents in a directory."""
    doc = OpenOCR(task='doc', use_layout_detection=True)

    os.makedirs(output_dir, exist_ok=True)

    for filename in os.listdir(input_dir):
        if filename.lower().endswith(('.jpg', '.png', '.pdf', '.bmp')):
            filepath = os.path.join(input_dir, filename)
            print(f"Processing: {filename}")

            result = doc(image_path=filepath)

            # Handle PDF (list) vs image (dict)
            if isinstance(result, list):
                for page_result in result:
                    doc.save_to_markdown(page_result, output_dir)
                    doc.save_to_json(page_result, output_dir)
            else:
                doc.save_to_markdown(result, output_dir)
                doc.save_to_json(result, output_dir)

    print(f"All results saved to {output_dir}")

process_documents('./docs', './output')

OCR with Custom Post-Processing

from openocr import OpenOCR
import re

def extract_structured_text(image_path, drop_score=0.5):
    """Extract and structure text from an image."""
    ocr = OpenOCR(task='ocr', drop_score=drop_score)
    results, _ = ocr(image_path=image_path)

    lines = []
    for result in results:
        for line in result:
            box = line[0]
            text = line[1][0]
            confidence = line[1][1]

            # Calculate bounding box center
            y_center = sum(p[1] for p in box) / 4

            lines.append({
                'text': text,
                'confidence': confidence,
                'y_center': y_center,
                'box': box,
            })

    # Sort by vertical position (top to bottom)
    lines.sort(key=lambda x: x['y_center'])

    return lines

result = extract_structured_text('page.jpg')
for line in result:
    print(f"{line['text']} ({line['confidence']:.2f})")

Formula Recognition

from openocr import OpenOCR

def recognize_formula(image_path):
    """Recognize mathematical formula from image."""
    unirec = OpenOCR(task='unirec')
    text, ids = unirec(image_path=image_path, max_length=2048)

    # UniRec outputs LaTeX for formulas
    print(f"LaTeX: {text}")
    return text

latex = recognize_formula('formula.png')
# Output: \frac{-b \pm \sqrt{b^2 - 4ac}}{2a}

Table Extraction

from openocr import OpenOCR

def extract_table(image_path):
    """Extract table content from image."""
    unirec = OpenOCR(task='unirec')
    text, ids = unirec(image_path=image_path, max_length=2048)

    # UniRec outputs LaTeX table format
    print(f"Table: {text}")
    return text

table_latex = extract_table('table.png')

Examples

Example 1: Batch OCR with Progress

from openocr import OpenOCR
import os

def batch_ocr(image_dir, output_dir='./ocr_results'):
    """OCR all images in a directory."""
    ocr = OpenOCR(task='ocr', mode='mobile')

    os.makedirs(output_dir, exist_ok=True)

    image_files = [
        f for f in os.listdir(image_dir)
        if f.lower().endswith(('.jpg', '.jpeg', '.png', '.bmp', '.tiff'))
    ]

    all_results = {}
    for i, filename in enumerate(image_files):
        filepath = os.path.join(image_dir, filename)
        print(f"[{i+1}/{len(image_files)}] Processing: {filename}")

        results, time_dicts = ocr(
            image_path=filepath,
            save_dir=output_dir,
            is_visualize=True,
        )

        texts = []
        for result in results:
            for line in result:
                texts.append(line[1][0])

        all_results[filename] = texts
        print(f"  Found {len(texts)} text lines")

    # Save all text
    with open(os.path.join(output_dir, 'all_text.txt'), 'w') as f:
        for filename, texts in all_results.items():
            f.write(f"--- {filename} ---\n")
            f.write('\n'.join(texts))
            f.write('\n\n')

    return all_results

results = batch_ocr('./images')

Example 2: Document to Markdown Converter

from openocr import OpenOCR
import os

def doc_to_markdown(input_path, output_dir='./markdown_output'):
    """Convert document images or PDFs to Markdown."""
    doc = OpenOCR(
        task='doc',
        use_layout_detection=True,
        use_chart_recognition=True,
    )

    os.makedirs(output_dir, exist_ok=True)

    result = doc(image_path=input_path)

    if isinstance(result, list):
        # PDF: multiple pages
        for page_result in result:
            doc.save_to_markdown(page_result, output_dir)
        print(f"Converted {len(result)} pages to Markdown")
    else:
        # Single image
        doc.save_to_markdown(result, output_dir)
        print("Converted image to Markdown")

    print(f"Output saved to: {output_dir}")

# Convert a scanned PDF
doc_to_markdown('paper.pdf')

# Convert a document image
doc_to_markdown('page.jpg')

Example 3: Multi-Task Comparison

from openocr import OpenOCR

def compare_tasks(image_path):
    """Compare results from different OpenOCR tasks."""

    # 1. Detection only
    det = OpenOCR(task='det')
    det_result = det(image_path=image_path)
    num_boxes = len(det_result[0]['boxes'])
    print(f"Detection: Found {num_boxes} text regions")

    # 2. End-to-End OCR
    ocr = OpenOCR(task='ocr')
    ocr_results, _ = ocr(image_path=image_path)
    ocr_texts = [line[1][0] for result in ocr_results for line in result]
    print(f"OCR: Extracted {len(ocr_texts)} text lines")
    for t in ocr_texts[:5]:
        print(f"  - {t}")

    # 3. Universal Recognition
    unirec = OpenOCR(task='unirec')
    text, _ = unirec(image_path=image_path)
    print(f"UniRec: {text[:200]}...")

    return {
        'det_boxes': num_boxes,
        'ocr_texts': ocr_texts,
        'unirec_text': text,
    }

compare_tasks('document.jpg')

Example 4: Gradio Demo Launch

from openocr import launch_openocr_demo, launch_unirec_demo, launch_opendoc_demo

# Launch OCR demo
launch_openocr_demo(share=True, server_port=7860, server_name='0.0.0.0')

# Launch UniRec demo
launch_unirec_demo(share=True, server_port=7861)

# Launch OpenDoc demo
launch_opendoc_demo(share=True, server_port=7862)

Limitations

• Text recognition accuracy depends on image quality
• Very small or heavily rotated text may reduce accuracy
• server mode requires PyTorch and is slower than mobile mode
• UniRec and Doc tasks use 0.1B parameter VLM, larger models may yield better results
• PDF processing converts pages to images internally, very large PDFs may use significant memory
• Complex handwritten text accuracy varies
• GPU recommended for best performance, especially for Doc and UniRec tasks

Installation

# Basic installation (CPU, ONNX backend)
pip install openocr-python

# GPU-accelerated ONNX inference
pip install openocr-python[onnx-gpu]

# PyTorch backend (for server mode)
pip install openocr-python[pytorch]

# Gradio demos
pip install openocr-python[gradio]

# All optional dependencies
pip install openocr-python[all]

# From source
git clone https://github.com/Topdu/OpenOCR.git
cd OpenOCR
python build_package.py
pip install ./build/dist/openocr_python-*.whl

Resources

• OpenOCR GitHub
• PyPI Package
• UniRec Paper
• OpenDoc Documentation
• Model Zoo & Configs