Facial Detection and Recognition for Security Applications

NOTE: This Repo and Readme is IN PROGRESS...

Problem

Facial recognition is challenging and yet a very useful technology if applied well. More information is coming....

Aim

In this project, we aimed to make a robust facial recognition system using state of the art models and methods to achieve high performance and solve the problem of spoofing.

Solution

ArcFace - anti-spoofing techniques

Data

The dataset contains headshot photos of people in the plain background from Adobe stock, and can be found in this Drive folder.(https://drive.google.com/drive/folders/1EYGouovWZR1JJCD4Yts7MnSMMDO3ODu4?usp=share_link) The test dataset contains headshot, half and full body shot photos of Chinese people in the various background from V3 dataset, and can be found in this Drive folder. (https://drive.google.com/drive/folders/1Y8Ykn1fGMD9NmBSSyFeTHqSvKEb7Ru0h?usp=share_link)

Checkpoints

The checkpoint file can be found in this Drive folder. (https://drive.google.com/drive/folders/1j0rRVoSuOvnCuP6bIaXP1XNnaaq_2sf4?usp=share_link)

Methodology

1. Preprocessing

   • Facial detection and cropping
   • Facial alignment
   • Uniform resizing

2. Verification : Anti-spoofing

   • At this stage, we will detect texture as either image or printed image. This helps us to differentiate the real person and the printed photo of the person.
   • In order to create a reasonable larger dataset for our model to learn, data augmentation was needed
     • For photo images
       • Horizontal and vertical flips
       • 90 degree and -90 degree rotations
     • For print images
       • Auto enhance
       • Horizontal and vertical flips
       • 90 degree and -90 degree rotations
   • LBP algorithm will be used to get feature vectors distinct to each texture
     • SVM or other models will learn features in order to classify whether this is an image or a printed paper
   • Hyperparameter tuning

3. Matching

   • Arcface: State of the art method that uses a special loss function called Additive Angular Margin penalty, it will do the facial recognition ArcFace head with ResNet backbone, and a function that detect the distance between two images so that it can classify if the person is registered or not.

Findings

Findings from the results will be added here.

Future Improvements

Next steps and the improvement points will be added to this session.

Repository Guide

This section will show the links of the related part of the project.