抄録
Autonomous observation and sampling of objects and organisms using autonomous underwater vehicles (AUVs) are desired. For this operation, it is first necessary to automatically identify the target from underwater images and other sources. Many of the previous studies employ supervised learning. However, due to insufficient underwater image data and its labels, supervised learning will not achieve sufficient performance. Furthermore, supervised learning cannot discover unseen targets, which are highly intriguing in scientific research. This study aims to extract distinctive objects in underwater images using a deep learning-based foreground segmentation method without additional training (without domain adaptation). This study specifically targets the extraction of aquatic animals and evaluates the performance using publicly available datasets for tracking marine animals.
As a result, this method successfully extracted marine animals to a certain extent without requiring any prior information or domain adaptation. This result suggests that the proposed method has the potential to be effective in detecting distinctive objects in underwater images.
This study will contribute to future autonomous underwater observation and sampling.
