Trends in Nearest Feature Classification for Face Recognition—Achievements and Perspectives
Mauricio Orozco‐Alzate, Csar Germn · 2009
We started this chapter with a critic about the overwhelming research efforts in face recognition, discussing benefits and drawbacks of such a situation. The particular case of studies related to linear dimensionality reduction is a per antonomasia example of an enormous concentration of attention in a small and already mature area. We agree that research topics regarding preprocessing, classification and even nonlinear dimensionality reduction are much more promising and susceptible of significant contributions than further studies in LDA. Afterwards, we reviewed the state of the art in prototype-based classification for face recognition. Several techniques and variants have been proposed since the early days of the nearest neighbor classifier. Indeed, an entire family of prototype-based methods arose; some of the family members are entirely new ideas, others are modifications or hybrid methods. In brief, three approaches in prototype-based classification can be distinguished: modifications of the distance measure, prototype generation and prototype selection methods. The last two approaches present dichotomies as shown in Fig. 1. We focus our discussion on nearest feature classifiers and their improved versions as well as on their use in dissimilarity-based classification. The main advantage of the RNFLS classifier is its property of generating feature line segments that are more concentrated in distribution than the original feature points. In addition, RNFLS corrects the interpolation and extrapolation inaccuracies of the k-NFL classifier, allowing us to use feature lines (in fact, feature line segments) in low dimensional classification problems. k-NFL was originally proposed and successfully used just in high dimensional representations such pixel-based representations for face recognition; however, thanks to the improvement provided by RNFLS, feature line-based approaches are also applicable now to feature-based face recognition. The k-NFP classifier is computationally very expensive. G-NFP can reduce the effort to a manageable amount, just by using a simple two-stage process: a GA-based prototype selection followed by the original k-NFP algorithm. We presented and compare three different but related approaches to generalize dissimilarity representations by using HMMs, clustering techniques and feature lines/planes respectively. The first two are model-based extensions for a given dissimilarity matrix and lead, in general, to lower dimensional dissimilarity spaces. In contrast, our methodology produces high dimensional dissimilarity spaces and, consequently, proper prototype