Options
2014
Conference Paper
Titel
Evaluation of methods for automatic fish segmentation
Abstract
Bycatch in the fishery endangers whole ecological systems in worldwide waters. By identifying and counting the marine animals of the bycatch, fishing techniques could be improved which would reduce the bycatch in turn. To ensure good identification, an excellent segmentation of the fish in photos taken from the bycatch is important. Six different segmentation methods are tested, which are Otsu's method, the Watershed algorithm, Region Growing, K-Means, K-Means distance probability and Gaussian Mixture Models (GMM) distance probability, with the two latter ones calculating probabilities for each pixel of an image for belonging to the fore- or the background. The methods are compared with the F-Measurement, the harmonic mean of precision and recall, to evaluate their accuracy. The result of this work will show that the best method for the seven tested fish species is the automatic K-Means algorithm, which is easy to use in practice.
Author(s)