Unknown Face Recognition

Abstract

This paper presents a method of recognising and tracking previously unknown faces in real-time video using the 720-pixel MacBook Pro Facetime HD camera. To identify faces in an image a Histogram of Oriented Gradients (HOG) and Support Vector Machine (SVM) based method was used. Embeddings for each face were then generated using a modified ResNet network. These embeddings were then compared using the Euclidean distance and if the distance was below a set threshold of 0.6, then the two embeddings were considered a match. If no match was found, then the embedding was recorded as a new face and used to recognise that face in later images. This method was tested on the Yale Faces dataset and had an accuracy of 98.18 %. In order to improve the runtime of the method, the image was reduced in size. The optimal image reduction was 60 % of the original size as this had no significant effect on the accuracy of the method but was able to achieve 7.7 fps with no display.

Face Recognition Setup

Instructions for installing the Face Recognition library can be found here.

Running the realtime detector

From the face_rec directory, execute:

python recognition.py

Optionally, the following parameters can be used to adjust the results:

• -s (int) Percentage of original size to process the image at (default at 50)
• -d (int) Percentage of original size to display the image at (default at 50)
• -f (directory) directory where pre-labelled images of faces are (default at "faces")

Running the single image detector

From the face_rec directory execute:

python recognition.py -i [path to image]

An example image has been provided, simply use "test.jpg" as the path:

python recognition.py -i test.jpg

Optionally, the following parameters can be used to adjust the results:

• -s (int) Percentage of original size to process and display the image at (default at 50)
• -f (directory) directory where pre-labelled images of faces are (default at "faces")

Evaluate the accuracy using Yale Faces dataset

From the face_rec directory execute:

python recognition.py -t [path to test directory]

Optionally, the following parameters can be used to adjust the results:

• -r (string) Image type fromn Yale Faces dataset to be used as the reference (default at _normal)

Options include:

• _normal
• _sad
• _happy
• _sleepy
• _surprised
• _wink
• _glasses
• _noglasses
• _rightlight
• _leftlight
• _centerlight