Environmental Microorganism Classification Using Sparse Coding and Weakly Supervised Learning

Abstract

Environmental Microorganisms (EMs), such as Epistylis and Rotifera, are very tiny living beings in human environments and decompose pollutants as their nutrition. The classification of EMs plays a fundamental role for establishing sustainable ecosystem. However, this is traditionally done by molecular methods using DNA or RNA analysis, or morphological methods using microscopes. These methods require huge monetary costs or manual efforts. In this paper, we propose an EM classification method by directly processing microscopic images. We especially address two problems: The first is a small training dataset problem. Because EM samples are collected from unstable real-world environments, it is difficult to stably obtain a large number of EM images. To effectively represent scarce training images, we use Sparse Coding (SC) which extracts sufficient local features from an image and reconstructs it by a sparse linear combination of bases. The second is a noisy image problem. Because EM samples are often found in dirty regions, their images often show a noisy background. To overcome this, we use Weakly Supervised Learning (WSL) which jointly performs the localisation and classification of EMs by examining the local information in training images. Experimental results on real-world images show the effectiveness and potential of our system.