Biologically-Inspired Identification of Plankton Based on Hierarchical Shape Semantics Modeling

This paper describes a novel hierarchical framework for automatic identification of plankton images, which is motivated by the semantics description of planktons used in the biology textbooks. The framework discretizes the identification of plankton into the recognition of various high-level shape semantics features. The semantics features are modeled with some manual instructions. Distinct from the previous approaches, such as "PCA+SVM" and "classifier stacking", our algorithm is more similar to the recognition procedure used by the biology experts, and the extracted features are more efficient for identification. The approach is tested on a collection of more than 2000 plankton images. Results demonstrate that the proposed approach has a satisfying classification accuracy and robustness to different number of training samples.