Dynamic classifier selection based on multiple classifier behaviour

Multiple classifier systems (MCSs) based on the combination of a set of different classifiers are currently used to achieve high pattern-recognition performances [1]. For each pattern, the classification process is performed in parallel by different classifiers and the results are then combined according to some decision “fusion” method (e.g., the majority-voting rule) [1]. The majority of such combination methods are based on the assumption that different classifiers make "independent" errors [1]. However, in real pattern recognition applications, it is difficult to design a set of classifiers that should satisfy such an assumption [1-5]. In order to avoid the errorindependence assumption, Huang and Suen proposed a combination method, named "Behaviour Knowledge Space" (BKS), based on the concept of multiple classifier behaviour (MCB) [2]. For each pattern, a vector whose elements are the decisions taken by the individual classifiers represents the behaviour of the MCS for such a pattern (see Section 2.2). In order to classify an unknown test pattern, all the training patterns exhibiting the same MCB of the test pattern are first identified. The classifications of such training patterns are then analysed, and the test pattern is assigned to the most frequent data class [2]. Another approach proposed to avoid the errorindependence assumption is the so-called “dynamic classifier selection” (DCS) [3-5]. DCS methods are aimed to select for each test pattern the classifier that will most likely classify it correctly. In this paper, a DCS method using MCB is proposed. It is worth remarking from the start that our paper is basically different from the work of Huang and Suen [2]. Our method exploits the concept of MCB for DCS purposes, while the BKS method is aimed at classifier combination. The DCS method we propose is based on the concepts of “classifier’s local accuracy” (CLA) and MCB. In particular, we exploit MCB information to compute CLA. The basic idea is to estimate the accuracy of each classifier in a local region of the feature space surrounding an unknown test pattern, and then to select the classifier with the highest value of this local accuracy to classify the test pattern [3-5]. In order to define such a local region and compute CLAs, the k-nearest neighbours of the test pattern are first identified in the training, or validation, data. The knearest neighbours characterised by MCBs “similar” to the one of the unknown test pattern are then selected to compute CLAs and perform DCS. This method is described in detail in the next Section. Experimental results and comparisons are reported in Section 3.