Face Recognition 人脸识别

Deep Learning, Index Page

doing

Face Recognition

Detection -> Alignment(~= landmark localization) -> Recognition

-> Recognition:

“Deep Face Recognition - A Survey” 这篇论文介绍了人脸识别领域的大致样貌。内容基本如下：

  • Background Concepts and Terminology
  • Components of Face Recognition
    • Data Preprocessing
    • Deep Feature Extraction
      Network Architecture
      Loss Function
      Similarity Comparison
  • Databases of Face Recognition
  • Real-World Scenes
    • Cross-factor FR
    • Hetorogenous FR
    • Multiple (or single) media FR
    • FR in industry

详细见博文 论文 Deep Face Recognition - A Survey。

-> -> Deep Feature Extraction

Network Architecture

人脸识别主要有两种思路。

  • 一种是直接转换为图像分类任务，每一类对应一个人的多张照片，比较有代表性的方法有DeepFace、DeepID等。
  • 另一种则将识别转换为度量学习问题，通过特征学习使得来自同一个人的不同照片距离比较近、不同的人的照片距离比较远，比较有代表性的方法有DeepID2、FaceNet等。

选取一些详细分析：

Real-World Scenes

  • Cross-Pose Face Recognition
  • Cross-Age Face Recognition
  • Makeup Face Recognition
  • NIR-VIS Face Recognition
  • Low-Resolution Face Recognition
  • Photo-Sketch Face Recognition
  • Low-Shot Face Recognition
  • Set/Template-Based Face Recognition
  • Video Face Recognition

Industry Concerns

  • 3D Face Recognition
  • Face Anti-spoofing
  • Face Recognition for Mobile Devices

Ref

[1] https://tech.meituan.com/deep_learning_image_recognition.html
[2] https://arxiv.org/pdf/1804.06655.pdf