Facial Recognition Using an Ensemble Model of Dimension Reduction Techniques and Convolutional Neural Networks

초록

With increasing trends in technology, there is a huge volume of data that is being created. One field in technology in which high amount of data is consumed, is computer vision. Human beings are able to detect and recognize faces or objects with ease even with external conditions such as expressions, illuminations or viewing angle affecting the sight when compared to the machines. This is because of high dimensions of data associated with it. Data dimensionality is refered to as total number of variables being measured in every observation. This project aims to compare different applicable dimensionality reduction techniques suitable for facial recognition system and propose an ensemble model of such techniques that will help improving the accuracy of the model and gauge the performance by testing it with different datasets consisting of facial images with varying illuminations, complex backgrounds, and expressions. The proposed ensemble model extracts feature vectors from a hybrid of two dimensional reduction techniques – Principal Component Analysis (PCA) and Locally Linear Embedding (LLE), and pass them through dense Convolutional Neural Network (CNN) to predict faces on the Labelled Faces in the Wild (LFW) dataset. The model performs with a testing accuracy of 0.95 and a testing F1 score of 0.94. The proposed system involves a webapp developed using Flask that captures a live video stream from a webcam which is integrated with the proposed ensemble model that allows the system to predict the face.

키워드