Soft-decision hierarchical classification using SVM-type classifiers

We address both recognition of true classes and rejection of unseen false classes, as occurs in many realistic pattern recognition problems. We focus on object recognition in which there are a large number of classes, and each object has different distorted (different aspect views) versions. We advance a binary hierarchical classifier and produce analog outputs at each node, with values related to the class conditional probabilities. This yields a new soft-decision hierarchical classifier (hard decisions are not made at each node). The hierarchy is designed by our new weighted support vector k-means clustering algorithm, which selects the classes to be separated at each node in the hierarchy using support vector information in higher-order space. Use of our SVRDM (support vector representation and discrimination machine) classifiers at each node provides generalization and rejection ability. The soft-decision SVRDM output allows use of the probability for each class at each node this is shown to improve classification (for true classes) and rejection (for false classes) performance. We provide classification and rejection results on the COIL-100 (Columbia Object Imagery Library) database that allows a large-class recognition and rejection problem to be addressed. We also provide test results on two UCI machine learning repository data sets this is to show that our soft-decision hierarchical SVRDM classifier applies to other pattern recognition problems that do not involve image data and geometric distortions of the data in each class, and to problems whose features are not image pixels as in the COIL-100 database.