EffiSeaNet: Pioneering Lightweight Network for Underwater Salient Object Detection

Qingyao Wu, Zhenqi Fu, Hong Lin, Chenyu Ma, Xiaotong Tu, Xinghao Ding; Proceedings of the Asian Conference on Computer Vision (ACCV), 2024, pp. 1486-1501

Underwater salient object detection seeks to pinpoint the most vital elements in underwater environments, offering considerable promise for underwater exploration. Considering the preference for low-complexity algorithms in underwater applications to maximize overall system efficiency, this paper proposes EffiSeaNet, a lightweight network designed to provide an effective solution for salient object detection in underwater scenarios. On the one hand, EffiSeaNet incorporates a parameter-free image enhancement block to mitigate the effects of image degradation caused by water. This block effectively addresses issues such as color distortion and reduced visibility, which are common challenges in underwater environments. On the other hand, we develop a customized lightweight network structure incorporating a novel cross-layer fusion strategy to efficiently capture and merge features. This enhances the network's ability to handle the variability and complexity of underwater objects and scenes while maintaining a low computational load. Extensive experiments on three public datasets demonstrate that our innovative designs achieve remarkable results while maintaining a low model size and computational complexity. This efficiency and effectiveness make our approach highly suitable for practical underwater applications where resources are limited, yet high precision is essential. Our code and results will be accessible to the public.