Hierarchical learning of tree classifiers for large-scale plant species identification

A hierarchical learning algorithm is developed for supporting large-scale plant species identification. A visual tree is first constructed for organizing large numbers of plant species hierarchically in a coarse-to-fine fashion. For the fine-grained plant species at the sibling leaf nodes under the same parent node, they share significant common visual properties but still contain subtle visual differences, a multi-task structural learning algorithm is developed to train their inter-related classifiers jointly for enhancing their discrimination power. For the coarse-grained categories at the sibling non-leaf nodes under the same parent node, a hierarchical classifier training algorithm is developed to leverage both the tree structure (i.e., inter-level constraint) and the common prediction structures shared among their sibling child nodes (i.e., inter-level visual correlation) to train their inter-related classifiers hierarchically. Our experimental results on large-scale plant images have demonstrated the effectiveness of our algorithm on large-scale plant species identification.