Blurriness-Guided Underwater Salient Object Detection and Data Augmentation
Yan‐Tsung Peng, Yu‐Cheng Lin, Wen‐Yi Peng, C.-L. Liu
Abstract
Salient object detection (SOD) has made significant progress with the help of deep networks. However, most works focus on terrestrial scenes, but underwater scenes for SOD are still little explored, which is essential for artificial-intelligence-driven underwater scene analysis. In the article, we propose and discuss two practical approaches to boost the performance of underwater SOD based on an inherent property of underwater scenes—blurriness, since an object appears more blurred when it is farther away. First, we utilize a self-derived blurriness cue and fuse it with the input image to help boost SOD accuracy. Next, we propose a blurriness-assisted data augmentation method that works for any available SOD model, called FocusAugment, for underwater SOD. We adjust images to enlarge differences between more- and less-focused regions based on the blurriness maps to augment training data. The experimental results show that both approaches can significantly improve state-of-the-art SOD models' accuracy for underwater scenes.