A dynamic classifier selection and combination approach to image region labelling

In this paper we propose a 'bank of classifiers' approach to image region labelling and evaluate dynamic classifier selection and classifier combination approaches against a baseline approach that works with a single best classifier chosen using a validation set. In this analysis, image segmentation, feature extraction, and classification are treated as three separate steps of analysis. The classifiers used are each trained with a different texture feature representation of training images. The paper proposes a new knowledge-based predictive approach based on estimating the Mahalanobis distance between test sample feature values and the corresponding probability distribution function from training data that selectively triggers classifiers. This approach is shown to perform better than probability-based classifier combination (all classifiers are triggered but their decisions are fused with combination rules), and single classifier, respectively, based on classification rates and confusion matrices. The experiments are performed on the natural scene analysis application.