OCR Locally

[macOS only] Use this skill when the user requests OCR (Optical Character Recognition), image/PDF text extraction. Uses macOS native Vision/PDFKit frameworks. Triggers: '识别图片', 'OCR', '提取图片文字', '提取PDF文字', '识别PDF', 'extract text from image', 'PDF OCR'.

Local OCR (macOS Only)

Overview

⚠️ Platform Requirement: This skill is macOS only. It requires macOS 10.15+ (Catalina or later) and uses macOS native frameworks:

• Vision framework - For OCR text recognition
• PDFKit framework - For PDF processing
• Core Graphics - For image rendering

This skill provides OCR (Optical Character Recognition) capabilities using macOS native Vision framework. It extracts text from images and PDFs without requiring any third-party libraries or internet connection.

Platform Requirements

⚠️ macOS Only - This skill cannot run on Linux, Windows, or other operating systems.

Required:

• macOS 10.15+ (Catalina or later)
• Vision framework (pre-installed on macOS)
• PDFKit framework (pre-installed on macOS)

Why macOS Only?

• Uses Vision framework for OCR (macOS/iOS only)
• Uses PDFKit framework for PDF processing (macOS/iOS only)
• Uses AppKit/Core Graphics for image handling (macOS only)

When to Use This Skill

Trigger this skill when the user:

• Requests OCR or image text extraction
• Mentions extracting text from images, screenshots, PDF files, or scanned documents
• Uses keywords like: "识别图片", "OCR", "提取文字", "提取PDF文字", "识别PDF", "extract text from image", "PDF OCR"
• Provides an image file or PDF file and asks to read or extract its content

Core Capabilities

1. Text Extraction from Images

Use scripts/ocr_vision_pro.swift for comprehensive OCR with the following features:

• Multi-language support (Chinese, English, Japanese, Korean, and more)
• Two output modes (mutually exclusive):
  • Text Mode (-t): Output only extracted text (default)
  • JSON Mode (-j): Output complete raw info including text, position, and confidence as JSON
• Confidence scores for each detected text block
• Bounding box information (text position in image)
• Output to console or file
• Precise or fast recognition modes

Basic usage:

bash

swift scripts/ocr_vision_pro.swift <image_path>

With options:

bash

swift scripts/ocr_vision_pro.swift <image_path> -l zh-Hans,en -o output.txt -f

2. Text Extraction from PDF Files

Use scripts/pdf_ocr.swift to extract text from PDF files with the following features:

• Extract text from specific pages or all pages
• Support page range specification (e.g., 1-5, 1,3,5)
• Two output modes (mutually exclusive):
  • Text Mode (-t): Output only extracted text (default)
  • JSON Mode (-j): Output complete raw info as JSON
• Same multi-language support as image OCR
• Precise or fast recognition modes

Basic usage (all pages):

bash

swift scripts/pdf_ocr.swift <pdf_path>

With page specification:

bash

# Single page
swift scripts/pdf_ocr.swift document.pdf -p 1

# Multiple pages
swift scripts/pdf_ocr.swift document.pdf -p 1,3,5

# Page range
swift scripts/pdf_ocr.swift document.pdf -p 1-5

# JSON mode
swift scripts/pdf_ocr.swift document.pdf -p 1 -j

3. Output Modes (Mutually Exclusive)

The script supports two output modes that cannot be used simultaneously:

Text Mode (Default, -t)

Outputs only the extracted text:

• Console output: Pure text
• File output (-o or -t): Saves text to file, optionally with separate confidence file

JSON Mode (-j)

Outputs complete raw information as JSON:

• Contains: image path, total blocks, average confidence, and per-block details
• Per-block info: index, text, confidence, bounding box (x, y, width, height)
• Outputs to stdout only (no file output options in JSON mode)

JSON output structure:

json

{
  "imagePath": "/path/to/image.png",
  "totalBlocks": 25,
  "averageConfidence": 0.85,
  "blocks": [
    {
      "index": 1,
      "text": "recognized text",
      "confidence": 0.95,
      "boundingBox": {
        "x": 0.10,
        "y": 0.20,
        "width": 0.30,
        "height": 0.05
      }
    }
  ]
}

4. Supported File Formats

Image Formats (ocr_vision_pro.swift):

• PNG (.png)
• JPEG (.jpg, .jpeg)
• TIFF (.tiff, .tif)
• BMP (.bmp)

PDF Format (pdf_ocr.swift):

• PDF (.pdf) - support single page, multiple pages, or page ranges
• Specify pages with -p option: 1, 1,3,5, or 1-5

4. Command-Line Options

Option	Description
-h, --help	Show help information
-t, --text	Text mode (default, output only extracted text)
-j, --json	JSON mode (output complete raw info as JSON)
-l, --language <lang>	Specify recognition language (comma-separated)
-o, --output <file>	Output text to file, auto-generate confidence file (<file>_confidence.txt)
-t, --text <file>	Output only complete text to specified file (text mode)
-c, --confidence <file>	Output only confidence details to specified file (text mode)
-f, --fast	Use fast mode (default: precise mode)

Note: -t (text mode) and -j (JSON mode) are mutually exclusive. JSON mode outputs to stdout only.

Supported languages:

• zh-Hans - Simplified Chinese
• zh-Hant - Traditional Chinese
• en - English
• ja - Japanese
• ko - Korean
• fr - French
• de - German
• es - Spanish
• it - Italian
• pt - Portuguese
• ru - Russian

Workflow

Step 1: Identify the Image Path

When the user requests OCR:

1. Ask for the image path if not provided
2. Accept common path formats: absolute paths, ~/path, or relative paths
3. Validate that the file exists before proceeding

Step 2: Determine Recognition Parameters

Based on user request or context:

1. Language: Default to zh-Hans,en for Chinese users, or en for English users
2. Mode: Use precise mode (default) for accuracy, fast mode (-f) for quick preview
3. Output: Ask if user wants results saved to file (-o option)

Step 3: Execute OCR

Run the OCR script with appropriate parameters:

bash

swift scripts/ocr_vision_pro.swift "<image_path>" -l zh-Hans,en

For saving to separate files (recommended):

bash

swift scripts/ocr_vision_pro.swift "<image_path>" -o "<output>"

This automatically creates two files:

• <output>.txt - Complete extracted text (pure text, no formatting)
• <output>_confidence.txt - Confidence details with statistics and per-block info

For separate text and confidence files with custom names:

bash

swift scripts/ocr_vision_pro.swift "<image_path>" -t "text.txt" -c "confidence.txt"

Step 4: Present Results

After OCR completes:

1. Display the extracted text to the user
2. If saved to file, inform the user of the output file path
3. Ask if user wants to:
   • Correct misrecognized characters
   • Process another image
   • Save results in a different format

Step 5: PDF Processing (if PDF file)

When processing a PDF file:

1. Identify PDF path and pages:

   • Ask for PDF path if not provided
   • Ask which pages to process (default: all pages)
   • Support formats: 1, 1,3,5, or 1-5

2. Determine recognition parameters:

   • Language: Default to zh-hans,zh-hant,en
   • Mode: Precise (default) or fast (-f)
   • Output: Text mode (default) or JSON mode (-j)

3. Execute PDF OCR:

bash

# All pages
swift scripts/pdf_ocr.swift "<pdf_path>"

# Specific pages
swift scripts/pdf_ocr.swift "<pdf_path>" -p 1,3,5

# Page range
swift scripts/pdf_ocr.swift "<pdf_path>" -p 1-5

# JSON mode
swift scripts/pdf_ocr.swift "<pdf_path>" -p 1 -j

4. Present results:
   • Text mode: Display text by page
   • JSON mode: Output complete JSON to stdout
   • Inform user of output format and options

Output Format

The script supports two mutually exclusive output modes:

Text Mode (Default, -t)

Outputs only the extracted text.

Console Output (without -o or -t <file>):

text

[Extracted text content]

File Output (-o option):

Creates <output>.txt with pure text.

With Confidence Details (console, when not using -j):

text

=== 置信度详情 ===

总识别块数: 25
平均置信度: 0.85

--- 逐块详情 ---

[1] Text content
    置信度: 0.95
    位置: x=0.10, y=0.20, w=0.30, h=0.05

--- 低置信度警告 (< 0.8) ---
[3] "Some text" - 置信度: 0.50

File Output with Confidence (-o option):

• <output>.txt - Complete extracted text
• <output>_confidence.txt - Confidence details

JSON Mode (-j)

Outputs complete raw information as JSON to stdout (no file output in JSON mode).

JSON Structure:

json

{
  "imagePath": "/path/to/image.png",
  "totalBlocks": 25,
  "averageConfidence": 0.85,
  "blocks": [
    {
      "index": 1,
      "text": "recognized text",
      "confidence": 0.95,
      "boundingBox": {
        "x": 0.10,
        "y": 0.20,
        "width": 0.30,
        "height": 0.05
      }
    }
  ]
}

Fields:

• imagePath: Path to the processed image
• totalBlocks: Total number of recognized text blocks
• averageConfidence: Average confidence score (0.0 - 1.0)
• blocks: Array of recognized text blocks
  • index: Block index (1-based)
  • text: Recognized text content
  • confidence: Confidence score (0.0 - 1.0)
  • boundingBox: Normalized bounding box coordinates (0.0 - 1.0)

Tips and Best Practices

1. macOS Only: This skill requires macOS. It will not work on Linux, Windows, or other operating systems.
2. Image Quality: Higher resolution images produce better results
3. Language Specification: Always specify language for better accuracy
4. Precise vs Fast: Use precise mode for final results, fast mode for testing
5. Batch Processing: For multiple images, use shell loop:

   bash

   for img in *.png; do
       swift scripts/ocr_vision_pro.swift "$img" -o "${img%.png}.txt"
   done
   

6. Confidence Threshold: Results with confidence < 0.5 may need manual verification

References

For detailed usage instructions and examples, load references/usage.md.

Resources

scripts/

• ocr_vision_pro.swift - Enhanced OCR script for images with full feature support
• ocr_vision.swift - Basic OCR script for simple use cases
• pdf_ocr.swift - PDF OCR script for extracting text from PDF files

references/

• usage.md - Comprehensive usage guide with examples