CE$^{3}$USOD: Channel-Enhanced, Efficient, and Effective Network for Underwater Salient Object Detection

Abstract

Underwater salient object detection (USOD) aims to identify the most crucial elements in underwater environments, holding significant potential for underwater exploration. Existing methods often overlook light degradation or involve larger network sizes, which are unsuitable for underwater mobile platforms and pose challenges to implement in practice. Given the importance of low-complexity algorithms in underwater applications to optimize system efficiency, this article introduces CE$^{3}$ USOD—an efficient network tailored to deliver an effective solution for salient object detection in underwater scenarios. On the one hand, we reconsider long-range dependencies and feature computation from a neighborhood perspective, leading to the development of the long-range context-aware module. Specifically, we approximate local and global context awareness by incorporating the maximum and average values of neighboring pixels within varying window sizes, which allows our method to achieve high performance while maintaining low computational cost. On the other hand, light scattering and absorption during underwater imaging frequently result in channel intensity imbalances in captured underwater images. To address this, we propose the color-guided pyramid aggregation module, which utilizes the weaker color channels enhanced by underwater image enhancement techniques as guiders for multiscale feature fusion, finally facilitating the model to obtain underwater domain information. Extensive experiments on four public benchmarks demonstrate that our innovative network achieves state-of-the-art results while maintaining a low model size (Params of 0.546M) and computational complexity (FLOPs of 0.416G). Therefore, CE$^{3}$ USOD proves to be effective and efficient, establishing its practicality, particularly for underwater applications.