Deep Learning based Detection, Segmentation and Counting of Benthic Megafauna in Unconstrained Underwater Environments

Abstract

Assessing and monitoring benthic communities is increasingly important in view of global alteration of marine environments. Deep learning has proven to effectively detect marine specimen in underwater imagery but still face problems with small input datasets, unconstrained environments and class imbalance. This study evaluates a data augmentation strategy to alleviate these limitations. Through synthetically derived image compositions, the entire input dataset was greatly extended from 700 to 12700 images. Additionally, specimen numbers of brittle stars, soft corals and glass sponges are equalized resulting in a mean average precision increase of 24 %. The overall mean average precision for box detections yields 76.7 and for instance segmentation 67.7 at an intersection over union threshold of 0.5. This study shows that deep architectures such as the deployed CenterMask via ResNeXt-101 model can successfully be trained with few original images from varying underwater scenes.