A Dynamic Programming Technique for Optimizing Dissimilarity-Based Classifiers

The aim of this paper is to present a dissimilarity measure strategy by which a new philosophy for pattern classification pertaining to dissimilarity-based classifiers (DBCs) can be efficiently implemented. Proposed by Duin and his co-authors, DBCs are a way of defining classifiers among classes; they are not based on the feature measurements of individual patterns, but rather on a suitable dissimilarity measure among the patterns. The problem with this strategy is that we need to measure the inter-pattern dissimilarities for all the training samples to ensure there is no zero distance between objects of different classes. Consequently, the classes do not overlap, and therefore, the lower error bound is zero. In image classification tasks, such as face recognition, one of the most intractable problems is the distortion and lack of information caused by the differences in face directions and sizes. To overcome the above problem, in this paper, we propose a new method of measuring the dissimilarity distance between two images of an object when the images have different directions and sizes and there is no direct feature correspondence. In the proposed method, a dynamic programming technique, such as dynamic time warping, is used to overcome the limitation of one-to-one mapping. Furthermore, when determining the matching templates of two images in dynamic time warping, we use a correlation coefficient-based method. With this method, we can find an optimal warping path by surveying the images in a one-dimensional or two-dimensional way (that is, with vertical-only scanning or vertical-horizontal scanning). Our experimental results demonstrate that the proposed mechanism can improve the classification accuracy of conventional approaches for an artificial data set and two real-life benchmark databases.