Self-Supervised Marine Organism Detection From Underwater Images

Abstract

In recent years, in light of the significant progress in deep learning on general object detection, research on marine organism detection has become increasingly popular. However, manual annotation of marine organism images usually requires specialized expertise, resulting in a scarcity of labeled data for research purposes. In addition, the complex and dynamic marine environment leads to varying degrees of light absorption and scattering, causing severe degradation issues in the collected images. These factors hinder the acquisition of high-quality representations for subsequent detection objectives. To overcome the reliance on annotated marine data sets and derive high-quality representations from extensive unlabeled and degraded data, we propose a self-supervised marine organism detection (SMOD) framework. To the best of the authors' knowledge, it is the first time that self-supervised learning has been introduced into the task of marine organism object detection. Specifically, in order to improve the quality of learned image representation from degraded data, a set of underwater augmentation strategies to improve the perceptional quality of underwater images is designed. To further address the challenging issue posed by numerous marine objects and diverse backgrounds, an underwater attention module is elaborately devised such that the model prioritizes objects over backgrounds during representation learning. Experimental results on URPC2021 data set show that our SMOD achieves competitive performance in the marine organism object detection task.