Face Detection and Recognition: ML Approaches

Among other tasks made possible through machine learning algorithms, face detection and recognition is a crucial computer vision task. To begin with, both face detection and recognition are co-related yet colloquially different. Face detection is a wider aspect than face recognition and is applied with the help of machine learning. Whether it is about face detection in surveillance, mapping images for medical diagnosis, or deep analysis of human faces in videos for intelligence purposes, ML models apply for both detection and recognition purposes.

Evolution: Machine learning approaches for face detection & recognition

Face detection is considered a part of the computer vision domain with relation to AI or Artificial Intelligence. Depending upon the dynamic nature of the face, studies have led to application of multiple face detection ML approaches such as Viola Jones algorithm, Kohonen approach, Local Binary Pattern Histogram (LBPH), Eigenfaces, Karhunen-Loeve detector. These approaches have helped in refining machine learning algorithms, which are today used in many face detection softwares and combined with ML models to provide accurate results. Let us have a look at them.

Eigenfaces: One of the earliest face recognition approaches includes Eigenfaces. When studies into the techniques for face detection started brewing, the Eigenfaces method was adopted. The Eigenfaces approach to face detection and recognition was based on the 2D perspective. It applied the covariance matrix based on probability distribution on a high dimensional vector space to produce Eigenvectors, basis of which face variations were produced for training. Later on, the approach was used with PCA or principal component analysis (a mathematical calculation), which defines points (for analysis) pertaining to the principal components of the data. The Eigenface approach made face detection little challenging due to its sensitivity to facial changes, lighting and environment.


Kohonen approach: Among early pioneers in face detection, the Kohonen approach also called the Self-Organizing Map (SOM) uses a network of Eigenfaces in a single-layered neural net with approximation of Eigenvectors through a serial correlation based matrix of facial images. The Kohonen approach is an unsupervised training method and often one or either two-dimensional. The approach, however, was less impactful in real-life environments. In the next stage, the Kohonen approach was augmented by using algebra for the recognition of faces.

Fisherface: Based on the reduction of face space dimension, and leveraging PCA or Principle Component Analysis that helps in reduction of dimensionality and focuses on projection direction maximizing the scatter of all classes of faces and images, Fisherface was developed by Robert Fisher for taxonomic classification. Fisherface algorithm uses LDA or Linear Discriminant Analysis for reducing dimensionality. Unlike Eigenfaces, the Fisherface method is able to handle facial variations, lighting and environment issues.

Viola Jones: With such an approach the Viola Jones method was adopted in 2001. It was a clutter-breaking approach, and the methodology applied in Viola-Jones framework revolved around detection of a face then extracting some features, followed by feature comparison in the existing data. It is based on binary classification of weak detectors. 
The ML algorithms for face detection and recognition in application

Face detection and recognition has always been an intriguing technology; applied through AI, combining machine learning methodologies to build a model that works. Coming a long way from following the traditional approaches for detection of human face and its features, the earliest of machine learning algorithms have strived to find human eyes first, followed by other features such as mouth, nose, eyebrows, and nostrils. Once a face is detected, the algorithm for face recognition gets down to work.

Post a Comment

0 Comments