README.md

metadata

title: Assistive Diagnostic Framework for Copyright
emoji: ⚖️
colorFrom: gray
colorTo: blue
sdk: gradio
sdk_version: 6.14.0
python_version: 3.10.14
app_file: app.py
pinned: false
license: mit

Assistive Diagnostic Framework for Copyright Infringement

This prototype application is an "Explainable AI" dashboard designed to assist in evaluating potential visual copyright infringement. Rather than outputting a single, opaque similarity score, this tool processes two images through a multi-model computer vision pipeline to map technical parameters to established legal criteria.

🧠 The Architecture

The pipeline divides the visual comparison into three distinct legal dimensions, optimized to run within standard memory constraints (e.g., Hugging Face Free Tier).

1. Semantic Match (The "Idea" Filter)

   • Model: CLIP (openai/clip-vit-base-patch32)
   • Function: Compares the overarching semantic concept of the images. This acts as a threshold mechanism to determine if the images share the same unprotected subject matter or "referent" before analyzing specific expressions.

2. Structural Layout (Substantial Similarity)

   • Model: OpenCV Canny Edge Detection
   • Function: Strips away style, texture, and color to compare only the fundamental structural outlines. Calculates the Intersection over Union (IoU) of the edge pixels to assess compositional overlap.

3. Patch Match (Fragmented Literal Similarity)

   • Model: DINOv2 (facebook/dinov2-base)
   • Function: Identifies "scattered literal copying." By extracting and normalizing local patch features, the model uses a Mutual Nearest Neighbors algorithm to map identical or near-identical fragments between the two images, regardless of their spatial location.

🚀 Running Locally

To run this application on your local machine, ensure you have Python 3.9+ installed.

1. Clone the repository:

   git clone <your-repo-url>
   cd <your-repo-directory>