Abstract: The endangered Acropora cervicornis coral has been considered a bellwether for coral reef habitat
change. Given the recent discovery of a concentrated region of growth off Belize, our group is studying
automatic abundance estimation of live A. cervicornis with the goal of future rapid assessment monitoring. In
this paper we present a novel technique for the automatic segmentation of coral image sets. While others have
had limited success applying machine learning techniques on color or texture-based features, our project
presents several confounding factors in acquisition and in image content for which we must compensate. Our
technique uses color features called quantile functions and SIFT texture features, and classifies local image
regions as either live A. cervicornis or other image content using linear Support Vector Machines. We present
promising results on a series of images for which we have manual segmentations to train with or test against.
We also compare our results to those achieved using established raw color features. Our approach may not only
greatly reduce the time cost of future abundance estimates of A. cervicornis, but also may be generalized to
other coral vision problems.
pdf of the paper
talk slides (pptx,
pdf)
mpg video showing traversal of the precision/recall curve
as the SVM threshold is relaxed.
ICRS Proceedings website