Considerations on extended feature vectors in automatic face recognition

Clearly, automatic recognition in large face populations will require many measurements. There have been few approaches which aim to generate such extended feature vectors. One approach considered combining several different sets, including feature descriptions and transform components, with apparent advantage accrued by the orthogonality of the measurements. More robust measures have included a new technique for eye location which employs concentricity using only few parameters and requiring little a priori information concerning a face's location. Further, a dual contour employing global energy minimization, again requires few parameters to provide measurements describing the face's boundary, again aimed at inclusion within an extended feature vector. Naturally, we seek to capitalize on minimal statistical correlation to improve recognition capability. To this end, we consider further the analysis of potential advantages of orthogonality, and show how this can indeed improve recognition capability. Accordingly, there is much research potential in extending the feature vector for automatic face recognition: there are rich avenues for future research in generation and combination of feature vectors for use in large face populations