A Self-learning Approach for Beggiatoa Coverage Estimation in Aquaculture

Beggiatoa is a bacterium that is associated with anoxic conditions beneath salmon aquaculture pens. Assessing the percentage coverage on the seafloor from images taken beneath a site is often undertaken as part of the environmental monitoring process. Images are assessed manually by observers with experience in identifying Beggiatoa. This is a time-consuming process and results can vary significantly between observers. Manually labelling images in order to apply visual learning techniques is also time-consuming and expensive as deep learning relies on very large data sets for training. Image segmentation techniques can automatically annotate images to release human resources and improve assessment efficiency. This paper introduces a combination method using Otsu thresholding and Fully Convolutional Networks (FCN). The self-learning method can be used to estimate coverage and generate training and testing data set for deep learning algorithms. Results showed that this combination of methods had better performance than individual methods.