Research Seminar - April 11, 2003

Face Detection and Recognition

Abstract:

This talk consists of two parts

Part I: Face Recognition and Content-Based Image Retrieval Using Mutiresolutional Eigenfeatures

In this work multiresolutional eigenfeatures are used in face recognition and content-based image retrieval applications. Principal Component Analysis (PCA) and Multivariate Discriminant Analysis (MDA) are useful multivariate statistical analysis tools for the construction of a compact vector space. An angle-free pose-tolerant face recognition system is proposed and implemented. Time and cost complexity of the process is considerably reduced by recycling the intermediate outputs of the training stage owing to a P-transform technique. A P-transformer is also implemented using Hard Description Language (HDL) for real-time applications. A content-based image retrieval system is demonstrated using PCA and MDA along with a wavelet decomposition technique. The results are compared among various feature vectors.

Part II: Face Detection Using Principal Wavelets

In this project we investigate a fast and robust face detection method using principal wavelets. Principal wavelets result from the combination of wavelet decomposition with Principal Component Analysis (PCA) or Non-negative Matrix Factorization (NMF). NMF has been shown to be a useful multivariate data analysis tool. However, the NMF basis is not orthogonal due to its non-negative constraint. We have developed an orthonormal NMF to achieve better performance, representation and expansion to various NMF applications. A fast and robust face detection algorithm is proposed. Face maps are efficiently constructed using the wavelet decomposition. The results of PCA and ONMF are compared. ONMF shows better performance with higher accuracy over PCA on untrained face images.