Facial Recognition Systems

Introduction

Facial recognition is the fastest growing technology with unbound scope and implementations. From the home surveillance camera to the most advanced security systems implemented at high-security zones, today's facial recognition market is a trillion-dollar industry.

Authentication of users using biometric data stored at some point in time is currently of utmost importance in the modern environment be it the cyber-security, or the modern criminal domains.

Unconstrained face recognition remains a challenging problem due to infra-class variations such as occlusion, disguise, varying orientations, face expressions, and age-variations. This causes much damage to modern detection models.

Brief History

During 1964 and 1965, Bledsoe, along with Helen Chan and Charles Bisson, worked on using the computer to recognize human faces (Bledsoe 1966a, 1966b; Bledsoe and Chan 1965). It was revealed that the Bledsoe’s initial approach involved the manual marking of various landmarks on the face such as the eye centres, mouth, etc., and these were mathematically rotated by computer to compensate for pose variation.

This project was labelled man-machine because the human extracted the coordinates of a set of features from the photographs, which were then used by the computer for recognition. Using a graphics tablet (GRAFACON or RAND TABLET), the operator would extract the coordinates of features such as the centre of pupils, the inside corner of eyes, the outside corner of eyes, point of widows’ peak, and so on. From these augmented data a scan was made through the central database looking for matched features and histograms.

Techniques Used

One of the most common methods is the use of LBPH recognizers. LBPH recognizers work on the principles of “Haar Cascades” and “Feature Extraction”.

It takes a 3×3 window and moves it across one image. At each move (each local part of the picture), compare the pixel at the centre, with its surrounding pixels. Denote the neighbours with intensity value less than or equal to the centre pixel by 1 and the rest by 0.

After consuming the whole image data, a list of local binary patterns is created.

LBP to binary conversion

Now, the list of binary patterns is converted to decimals and a histogram is obtained. A sample histogram looks like this:

Histogram sample

The histogram obtained is unique for each label. The algorithm also keeps track of which histogram belongs to which person. Next, the recognizer model compares this intensity histogram to Haar features extracted from images when training present in the dataset.

Finally, we deduce confidence scores for each set of the image in the dataset and match with the best score.

Further Challenges and Improvements

One of the main challenges for any recognition system is the orientation of face captured and lighting conditions. The efficiency of any algorithm deteriorates in each case. But nowadays, with the use of CNN’s and availability of GPU’s the models can train more accurate and adapt to augmented data, when using a Neural Network.

One can check my facial recognition software made using refined-LBPH techniques at https://github.com/projjal1/Facial_Recognition_Software.