FaceDuplicationFilter

中文

Content

• FaceDuplicationFilter
  • Content
  • Project Overview
  • Technical Innovations
  • References
  • System Requirements
  • Environment Setup and Usage
    • Method 1: Complete Package
    • Method 2: Conda Installation
  • License

Project Overview

This project is based on the eDifFIQA model, combining its face quality assessment capabilities to develop a face quality-enhanced duplicate data filtering system for identifying and removing duplicate face images.

Face recognition technology has long been a hot topic in cutting-edge academic research. With the advancement of AI technology, various face recognition techniques heavily rely on large-scale training datasets. However, these datasets often contain significant redundant data which, if not properly filtered, may become high-frequency noise that affects model training. This could lead to suboptimal training results, slower convergence, or even abnormal gradient changes during training. Therefore, establishing a reliable face data cleaning system is crucial.

Our system addresses this need by utilizing a carefully trained face quality assessment model to effectively extract the most distinctive face data while filtering out low-distinctiveness redundant data. This process not only improves training efficiency but also significantly accelerates model convergence, thereby enhancing overall system performance and reliability.

Technical Innovations

1. K-Fold Cross Validation for Optimal Face Selection:

   • We employ k-fold cross validation to obtain optimal face combinations
   • The evaluation criteria is based on distances between faces
   • Particularly effective for our product's scenario of small-batch, high-duplication applications

2. Enhanced Face Representation with eDifFIQA:

   • Compared to traditional average distance methods, we introduce the eDifFIQA model for quality assessment
   • Use quality-weighted averages as face representation
   • Considers factors like face angle, noise, brightness, and camera distortion

3. Advanced Quality Assessment Methodology:

   • Diffusion process using a custom UNet model for generating noisy and reconstructed images
   • Process repeated on horizontally flipped images to capture pose impact
   • Quality score calculation through embedding comparison
   • Enhanced with knowledge distillation and label optimization:
     • Quality label optimization using relative position information from FR model embedding space
     • Representation consistency loss (Lrc) and quality loss (Lq) for improved prediction

References

@article{babnikTBIOM2024,
  title={{eDifFIQA: Towards Efficient Face Image Quality Assessment based on Denoising Diffusion Probabilistic Models}},
  author={Babnik, {\v{Z}}iga and Peer, Peter and {\v{S}}truc, Vitomir},
  journal={IEEE Transactions on Biometrics, Behavior, and Identity Science (TBIOM)},
  year={2024},
  publisher={IEEE}
}

System Requirements

• Windows environment (Python embedded package)
• Linux environment (theoretically supported, requires Python configuration)

Environment Setup and Usage

Method 1: Complete Package

1. Download the complete package via HuggingFace linkDownload
2. Extract the package
3. Run start.bat on Windows

Method 2: Conda Installation

1. Create a Python 3.10 environment
2. Install packages from requirements.txt in the project root
3. Download model weights according to instructions from eDifFIQA:
   • Recommended weights: r100.pth and ediffiqaL.pth
   • Place them in the weights folder
4. Run allmain.py

License

Follow the original project eDifFIQA, open source licenses.