An Active Learning Approach to Egeria Densa Detection in Digital Imagery

Abstract

This chapter focuses on the development of an active learning approach to an image mining problem for detecting Egeria densa (a Brazilian waterweed) in digital imagery. An effective way of automatic image classification is to employ learning systems. However, due to a large number of images, it is often impractical to manually create labeled data for supervised learning. On the other hand, classification systems generally require labeled data to carry out learning. In order to strike a balance between the difficulty of obtaining labeled images and the need for labeled data, we explore an active learning approach to image mining. The goal is to minimize the task of expert labeling of images: if labeling is necessary, only those important parts of an image will be presented to experts for labeling. The critical issues are: (1) how to determine what should be presented to experts; (2) how to minimize the number of those parts for labeling; and (3) after a small number of labeled instances are available, how to effectively learn a classifier and apply it to new images. We propose to use ensemble methods for active learning in Egeria detection. Our approach is to use the combined classifications of the ensemble of classifiers to reduce the number of uncertain instances in the image classification process and thus achieve reduced expert involvement in image labeling. We demonstrate the effectiveness of our proposed system via experiments using a real-world application of Egeria detection. Practical concerns in image mining using active learning are also addressed and discussed.2