An interdisciplinary team of computer scientists, mechanical engineers, and entomologists from Oregon State University and the University of Washington are developing computer vision, machine learning, and robotic methods for high-precision generic object recognition and applying these methods to the imaging and classification of invertebrate specimens of soil mesofauna and freshwater zooplankton. Current manual methods for recognizing and counting these organisms are extremely tedious and time-consuming, and require a high degree of expertise. Automated, rapid-throughput population counting will provide a revolutionary new tool for ecologists to understand and monitor soil and freshwater ecosystems. Soil arthropods form a central component of ecological processes in soils, so accurate soil arthopod population counting is critical to improving our understanding of ecosystem functions and community ecology. Freshwater zooplankton species are a fundamental component of many ecosystems, because they transfer energy from primary producers to consumers such as fish and birds. Zooplankton also serve as a model system for understanding basic ecosystem processes, predator-prey dynamics, and disease ecology.
Automated recognition of these organisms poses difficult classification problems because it requires much more precise discrimination than generic object recognition tasks of the type commonly studied in computer vision. Current approaches to generic object recognition employ a bag-of-keypoints methodology in which hand-crafted region detectors, hand-crafted region descriptors, and unsupervised feature dictionaries are applied to convert an image into a fixed-length feature vector. Machine learning is only employed at the final step to classify this feature vector into a generic object class. This project seeks to integrate machine learning into all aspects of the vision pipeline. It will develop and test discriminative learning algorithms for the automated discovery of region detectors, region descriptors, feature dictionaries, and classifiers. To reduce the risk of overfitting, sub-part correspondences and spatial constraints will be imposed to constrain the learning algorithms. In addition to discriminative methods, the investigators will also learn generative models to help reject debris and unknown species that appear in the images. Model adaptation methods will be developed to take advantage of the fact that in any given biological sample, organisms of the same taxon tend to be more similar to each other than they are when samples from multiple sites are pooled.
Progress on this project will be regularly reported at http:// web.engr.oregonstate.edu/~tgd/bugid/
