FACE VERIFICATION WITHOUT FALSE ACCEPTANCE

Abstract

Principal Components Analysis (PCA) and Linear Discriminant Analysis (LDA) are two popular approaches in face recognition and verification. The methods are classified under appearance-based approach and are considered to be highly-correlated. The last factor deems a fusion of both methods to be unfavorable. Nevertheless the authors will demonstrate a verification performance in which the fusion of both method produces an improved rate compared to individual performance. Tests are carried out on FERET (Facial Recognition Technology) database using a modified protocol. A major drawback in applying LDA is that it requires a large set of individual face images sample to extract the intra-class variations. In real life application data enrolment incurs costs such as human time and hardware setup. Tests are therefore conducted using virtual images and its performance and behaviour recorded as an option for multiple sample. The FERET database is chosen because it is widely used by researchers and published results are available for comparisons. Performance is presented as the rate of verification when false acceptance rate is zero, in other words, no impostors allowed. Initial results using fusion of two verification experts shows that a fusion of T-Zone LDA with Gabor LDA of whole face produces the best verification rate of 98.2% which is over 2% improvement compared with the best individual expert.