Image Reader

Extract text from images using OCR (Optical Character Recognition). Use this skill when you need to read text content from images, screenshots, photos, or any image file. Supports Chinese and English text with high accuracy. Triggers when user sends an image and asks to read/extract/identify text from it, or when you receive an image that contains text you need to understand.

Image Reader - OCR Text Extraction

A high-performance OCR skill for extracting text from images. Powered by RapidOCR with PP-OCRv4 models, supporting Chinese and English text recognition.

Features

• Multi-language: Chinese (simplified/traditional), English, and mixed text
• High accuracy: PP-OCRv4 model with >95% accuracy on typical screenshots
• Structured output: Text with confidence scores and bounding boxes
• Image info: Dimensions, format, and color mode included
• Fast: CPU-only, no GPU required

Quick Start

bash

python scripts/read_image.py /path/to/image.jpg

Usage Examples

Extract text from a screenshot

bash

python scripts/read_image.py screenshot.png

JSON Output

The script outputs structured JSON:

json

{
  "success": true,
  "text": "Full extracted text",
  "lines": [
    {
      "text": "Individual line",
      "confidence": 0.98,
      "box": [[x1,y1], [x2,y2], [x3,y3], [x4,y4]]
    }
  ],
  "line_count": 5,
  "image_info": {
    "format": "PNG",
    "size": [1920, 1080],
    "mode": "RGB"
  }
}

Requirements

bash

pip install rapidocr onnxruntime pillow

First run will download OCR models (~50MB) automatically.

Common Use Cases

• UI Screenshots: Extract text from app/website screenshots
• Document Photos: Read text from photographed documents
• Diagrams: Extract labels and annotations
• Receipts: Parse receipt/invoice data

Output Fields

Field	Type	Description
success	bool	Whether OCR succeeded
text	string	All extracted text
lines	array	Individual text lines with metadata
line_count	int	Number of text lines detected
image_info	object	Image metadata

Technical Details

• Engine: RapidOCR (ONNX Runtime backend)
• Models: PP-OCRv4 (detection + recognition)
• Languages: Chinese, English (auto-detected)
• Performance: ~1-2 seconds per image on CPU

License

MIT License

Third-party dependencies:

• RapidOCR - Apache 2.0 License
• ONNX Runtime - MIT License