Face Recognition Using Eigenfaces

Abstract

In this thesis, we researched various methods leading to recognition of a person's face by ways of statistical and mathematical analysis and comparisons of facial features to that of a known person. Many approaches to overcome inherent face recognition challenges have been developed over the years. One of the most accurate and rapid ways to identify faces is to use what is called the eigenface [1] technique, that was created as a linear combination model using the mathematical software called MatLab. The linear combination model has been recognized for several years and uses the process of Principal Components Analysis (PCA). This system was able to successfully recognize all randomly generated photos (mug-shots) with 97.7 percent accuracy using 21 eigenfaces. The eigenface technique uses a highly effective combination of linear algebra and statistical analysis (PCA) to generate an identifying set of base faces, the eigenfaces, against which the inputs are tested and matched. Although using a sophisticated statistical model is a means to recognize a person by facial patterns, it is also critically important to ac-knowledge that the collected data is imperfect and requires some manipulation to be both clean and normalized. The objective is to represent a face as a linear combination of images from our data base. Recently, Random Projection (RP) has emerged as a powerful method for dimensionality reduction. In this paper, I will compare and contrast Random Projection (RP) with PCA using a well known face database. The experimental results illustrate that although PCA represents faces in a low-dimensional subspace, the overall performance is comparable to that of Random Projection, having higher computational requirements and being data dependent.