Underwater Mussel Segmentation Using Smoothed Shape Descriptors with Random Forest

Segmentation of objects of interest is no longer a massive challenge with the adoption of machine learning and AI. However, feature selection and extraction are not trivial tasks in these approaches, and it is often necessary to introduce new methods for the creation of such features. Due to the lack of control over environmental conditions, for example turbidity and light scatter for underwater data, it is difficult to acquire color and texture features. However, it is still possible to obtain satisfactory shape features. This has led to the development of methods that can generate shape descriptors for use as features for data segmentation using machine learning methods such as random forest. In this work, we introduce a smoothed shape descriptor, which is the basis for a set of features used for the segmentation of underwater mussel structures with an accuracy of almost 90% based on manually labeled and measured mussel clusters by professional divers.