Reducing Annotation Times: Semantic Segmentation of Coral Reef Survey Images

Benthic quadrat studies requiring time-intensive manual image annotation are currently a critical component of assessing the health of coral reefs. Patch-based image classification using convolutional neural networks (CNNs) can automate this task by providing sparse labels, but remain computationally inefficient. This work extends the idea of automatic image annotation by using fully convolutional networks (FCNs) to provide dense labels through semantic segmentation. We present an improved version the Multilevel Superpixel Segmentation (MSS) algorithm, which repurposes existing sparse labels for images by converting them into the dense labels necessary for training a FCN automatically. Our implementation-Fast-MSS-is demonstrated to perform considerably faster than the original without sacrificing accuracy. To showcase the applicability to benthic ecologists, we validate this method using the Moorea Labeled Coral (MLC) dataset as a benchmark. FCNs are evaluated by comparing their predictions on test images with the corresponding ground-truth sparse labels. Our results indicate that FCNs' perform with accuracies that are suitable for many ecological applications, and can increase even further when trained on dense labels augmented with additional sparse labels provided by a patch-based image classifier. The contributions of this study help move the field of benthic ecology towards more efficient monitoring of coral reefs through entirely automated processes.