Plankton Image Classification Based on Multiple Segmentations

Abstract

Due to image quality related issues, classification of plankton images, particularly of those collected in situ, strongly relies on shape features. Thus, image segmentation is a critical step in the classification pipeline. In general, the segmentation algorithm that leads to the best overall classification accuracy does not necessarily imply best classification accuracy with respect to each of the individual classes. In addition, in real time applications, changes in the environment or in the image acquisition devices require fast adjustments in the classification pipeline. Customizing segmentation algorithms for each situation may demand considerable effort. Motivated by these issues, we address the problem of using multiple segmentation algorithms and letting the classifier decide how to make best use of them. Some case studies and results are presented and discussed.