Semantic Segmentation of seafloor images in Philippines based on semi-supervised learning

Semantic segmentation of marine images can be used to describe seafloor scenes and monitor marine creatures. However, preparing human-annotated datasets for image segmentation is time-consuming task. Therefore, this paper proposes a semi-supervised semantic segmentation algorithm based on the combination of Mean-Teacher and U-Net models to classify seafloor images collected in Philippines. The method will train and validate on two parts of the image. On the one hand, for images containing categories of coral, sea urchin, sea stars, and others (including sediment and seagrass), ordinary labeling is used for training and validation. On the other hand, for images only including seagrass and sediment categories, manual labeling of seagrass categories is particularly difficult. In order to overcome this barrier, based on the characteristics of this type of images, K-means clustering algorithm is used to obtain labeled dataset for training and validation. Compared with the U-Net based supervised method, the semi-supervised method proposed in this paper achieves good results and accuracy values even with fewer labeled images.