A hierarchical laplacian TWSVM using similarity clustering for leaf classification

This article introduces a multi-class hierarchical algorithm for semi-supervised classification. The proposed algorithm incorporates the benefits of tree-based classification approaches and vastly available unlabelled data. It also overcomes the deficiency of existing multi-class extension approaches viz . Non-linearity, imbalanced class classification and increasing classification cost with increasing the number of classes. The proposed algorithm selects classes from a pool and creates two clusters with a notion of maximum inter-cluster distance and intra-cluster similarity. The efficiency and effectiveness of the proposed algorithm are evaluated using one artificial dataset, three benchmark datasets viz. iris , wine , and seeds . The article presents an application of the algorithm for a real-world and complex plant recognition problem using three leaf image datasets i.e. Flavia , Swedish and self-collected . The experimental results confirm that the hierarchical extension serves several benefits, including efficient classification accuracy, less computational cost, faster classification, and reduced class imbalance. An improvement of 11% classification rate on self-collected leaf images with one-fourth of the computational cost required by tree-based laplacian TWSVM confirms its applicability on plant classification and other domains as well.