Seal Extractor — AI skill for Claude Code

Seal Extractor

**Extract red seals (公章/印章) from scanned documents with pixel-perfect fidelity.**

A Claude Code skill that removes black table lines, handwritten text, and annotations from scanned seals — preserving the original red ink pixels exactly as they are, with optional AI-powered gap repair.

The Problem

When you scan a document with a red official seal (公章), the seal overlaps with:

• Black table grid lines
• Handwritten annotations (dates, "盖章", signatures)
• Other printed text

You need the clean seal isolated — for archiving, reprinting, or digital use — but manual Photoshop work is tedious, and AI image generators change the font style.

The Solution

A two-step pipeline that combines the best of both worlds:

Step 1: OpenCV Color Filtering (local, fast, deterministic)

• HSV + Lab color space analysis detects red pixels with high precision
• Black/gray elements (lines, text) are detected and removed
• Morphological operations clean noise and fill small internal gaps
• Result: original seal pixels preserved exactly — zero AI hallucination

Step 2: AI Gap Repair (optional, via OpenRouter)

• Where black lines crossed over the red seal, small white gaps remain after Step 1
• Calls Gemini image models to naturally inpaint these gaps
• Preserves original font style and proportions
• Models tried in order: Gemini 3 Pro → 3.1 Flash → 2.5 Flash

Installation

claude skill install 24kchengYe/seal-extractor

Or clone manually:

cd ~/.claude/skills/
git clone https://github.com/24kchengYe/seal-extractor.git

Dependencies

pip install opencv-python numpy

Usage

In Claude Code

Just tell Claude what you need:

提取这个公章
extract the seal from this image
把公章抠出来，去掉黑线

Claude will automatically use this skill.

Command Line

**Basic extraction (transparent background):**

python scripts/extract_seal.py input.png -o seal_clean.png

**White background:**

python scripts/extr