Computer Vision — Facial Recognition using Gaussian Mixture Models, Adaboost & Haar Features, and more...

  ECE 763 (Computer Vision) Projects

  Individual Project (& one Group Project) — Status: Complete

  Jan 2021 - May 2021 — Raleigh, North Carolina

Overview:

1. Project 1: Facial Detection using Classical Machine Learning with Various Mixture Models
• 
• Mixture models are using some distribution in N-dimensional space and clustering for classification. For example, the figure above shows a data in 2 & 3-D spaces and how they would be classified using a Gaussian mixture model. Note that these classification models can be distinct (where each cluster is a seperate form of classification) or merged (where everything a cluster is something, and everything excluded is not something). For facial detection, a merged model is used in a high dimensional space for classifying whether something is a face or not. The dimensionality is based off the number of pixels in the training images. Below are 20x20 (row 1) and 25x25 (row 2 & 3) Gaussian mixture models with 8 clusters. The means of each cluster are displayed below:


• 


• For this project, a t-Distribution, mixture of t-Distribution, as well as factor analysis were completed. More can be read in the report below, see the Literature section.


2. Project 2: Facial Detection using Viola-Jones Algorithm: Adaboost + Haar-like Features
• Haar-like features can be used in junction with Adaboost to create a lightweight and fast facial detection model. It is colloquially known as the Viola-Jones object detection algorithm, but has been popularized for facial recognition. It is commonly used by digital cameras to detect faces prior to taking photographs. For this project, this was implemented and trained using a face dataset. To achieve the lowest of accuracies, only the first 10-20 features need to be checked, but to achieve the full accuracy of the model, all features must be checked in a cascade fashion. Below is a sample image alongside the 15 strongest features for the face dataset. A total of ~250 features were used in multiple models for this project. More can be read in the report below, see the Literature section.


• 


3. Project 3: Facial Detection using Simple Deep Learning
• Using a simple 2D CNN implemented using Python (Keras / Tensorflow), a facial detection model was made on the same dataset used in all 3 projects. Both training and validation accuracy were significantly higher than the previous two projects, which shows the power of deep learning, considering the fact that the model architecture is nothing novel, the dataset was not flawless, and the students in the class (including my teammates at the time) were relatively inexperienced with deep learning and its implementation. I would show visuals for this project as I did with the other two above, however all I would show are datapoints with their respective classification. A binary classification metric was used. More can be read in the report below, see the Literature section.

My Contribution:

For Project 3: Group Project — 3 members

The work for this project was more evenly spread out, as compared to previous projects I have had in graduate school. I did the data pre-processing from the face dataset, trained multiple models multiple datasets, visualized the intermittent layers of our most effective model, and wrote my fair share of the report. In can be noted in the report, that each member of the team had their own unique way of tackling the data pre-processing, so we trained our best selected models on all our datasets and included the results in the report.

Literature:

Media Citation(s):

Gaussian Mixture Model visual - https://stanford.edu/~cpiech/cs221/img/kmeans3d.png