Face Detection using PCA with People and Expression Classifiers

Real-time detection implementing:

• Principal Component Analysis (PCA)
• Nearest Neighbour Classifier (SSD)
• Support Vector Machines (SVM)

Video demos:

• Nearest Neighbour Classifier (SSD)
  • Extreme recognition of single person
    • https://youtu.be/bO-e7UspUXU
  • Recognition of people within and outside the recognised list. Tested using people found on Google Images
    • https://youtu.be/FJaFxk28X5Y
• Support Vector Machine Classifer (SVM)
  • Demo oriented around extreme recognition of single person
    • https://youtu.be/Zb_AllFWQ64
  • Recognition of people within and outside the recognised list. Tested using people found on Google Images
    • https://youtu.be/3laCd__XjS4

Docker Image

This app is available as a docker image, for fitting PCA and training classifiers. To execute:
docker pull foorenxiang/faceid
docker run -ti --rm faceid
cd ~/opencv/
python3 trainFace.py

Scripts to run:

• trainFace.py
  • Run this script to:
    • Fit the PCA model
    • Project faces in training images to eigenspace
    • Train/test person and expression classifiers
• rtFaceID.py
  • Run this script to perform real-time face detection
• videoFaceID.py
  • Run this script to perform face detection on a video file

Custom PCA Module

• rxPCA.py

Data pre-processing scripts:

• adjustDict.py
  • Drop eigen coordinates of people to be removed from training set (reduce number of person classes)
• adjustDict(Remove expressions).py
  • Drop eigen coordinates of certain expressions (remove invalid expression classes. e.g. denoting head rotation instead of actual expressions)
• centroidFinder.py
  • Calculate eigen coordinate centroid for updating old pcDict.bin files without centroid data
• faceCropper.py
  • Tool for cropping faces from images
• reSizer.py
  • Image resizer script
• rxPCAtest.py
  • Script for testing custom PCA module
• suffixPhotos.py
  • Script for adding suffix to photo names

Expression Analysis script

• expressionsAnalysis.q
  • Perform exploratory data analysis on eigen coordinates of expressions
  • Expression distribution can be plotted using Kx Developer

Classifiers Tested during training step:

• Nearest Neighbour Classifier (SSD)
• Support Vector Machines
• Adaboost
• GradientBoost
• XGBoost
• Stack Generalizer
• Random Forest
• Histogram Gradient Boosting
• K-NN
• Extreme Learning Machines

Images and video-stream are zero-meaned to deal with lighting changes.

SSD used to determine if person in frame is not within training set.

EE4208 Intelligent System Design Module Assignment 1 Group Coursework

Due to privacy concerns, the training dataset and generated models are not included in this repository.
To fit the PCA principal components and classification models, place your photos in a folder named input, with the notation: person_expression.jpg
E.g. ./input/john_smiling.jpg

Team Members:

• Calista Lee
• Chew Simin
• Choo Yaw Feng
• Foo Ren Xiang
• Zhang Zeyu

Libraries Used:

• Scikit-Learn
• OpenCV-Python
• Python-ELM by D. Lambert
  • https://github.com/dclambert/Python-ELM

Dataset Used:

• CAS-PEAL Face Database (neutral and expressions sets)
  • http://www.jdl.ac.cn/peal/
• Group Member Photos
  • 5 photos of each person with varying head tilt

429 images of 83 people with varying expressions were used to train expression classifiers
112 images of 22 people were used to train person classifiers