The computer does the hard work: (semi)-automatic classification of zooplankton

Manual enumeration and measurement of zooplankton has always been a bottleneck in plankton studies. With the advances in numeric imaging systems, both in situ and for the analysis of traditional plankton net samples in the laboratory, new opportunities appear to automate plankton analysis using computers. Whatever the initial images and whatever the purpose of the analyses, the scheme is always the same: (1) images are analyzed and particles are identified (segmented), (2) various features are extracted, (3) a small subsample is manually identified to make a training set, (4) a classifier is trained to recognize these particles, and (5) that classifier is finally used to identify all plankters in the series. Even if this scheme is well-known and successfully applied to various machine vision applications (for instance, face recognition, car numberplate identification), its application to plankton is challenging due to the extreme variability of planktonic critters. Several recent studies show, however, that current algorithms can separate 20-30 taxa or more, with a success level of about 75-80%. This is enough for many ecological studies, and a good starting point for other applications. Given the common framework and current performances of such systems, we propose a free software, called ZooImage (http://www.sciviews.org/zooimage/), tailored for such analyses. It is both a highly customizable toolbox to build actual applications, and a complete environment ready to routinely analyze plankton digital images. Its mode of operation ranges from completely automatic analysis to semi- automatic support for taxonomists and operators working with plankton digital images. It is not a proprietary system, but a project aiming to be developed collaboratively by the scientific community (e.g., the 'RAPID' SCOR working group for automatic visual plankton identification). The main benefit expected from such a collaboration is a coherent environment to analyze numeric plankton images whatever their origin, and without unnecessary duplication of the development effort across separate and incompatible projects.