Heterogeneous Experts and Hierarchical Perception for Underwater Salient Object Detection

Area of Science:

• Computer Vision
• Image Processing
• Artificial Intelligence

Background:

• Existing underwater salient object detection (USOD) methods often use fusion strategies but neglect the distinct characteristics of different data modalities (e.g., RGB and depth).
• This limitation hinders the effective integration of multimodal information for accurate object detection in underwater environments.

Purpose of the Study:

• To develop a novel network, the Heterogeneous Experts and Hierarchical Perception network (HEHP), for improved underwater salient object detection.
• To address modal discrepancies and enhance the learning of disentangled representations from RGB and depth data.

Main Methods:

• Proposing Hierarchical Prototype guided Interaction (HPI) for fine-grained alignment and refinement using complementary modalities.
• Introducing Mixture of Frequency Experts (MoFE) and Four-Way Fusion Experts (FFE) to model and integrate hierarchical spatial and frequency information.
• Implementing Uncertainty Injection (UI) to manage noise in depth maps and Holistic Prototype Contrastive (HPC) loss for robust representation learning.

Main Results:

• The HEHP network demonstrated superior performance compared to state-of-the-art binary detection models on two USOD datasets and four underwater scene benchmarks.
• The method achieved impressive results on seven natural scene benchmarks, highlighting its scalability and generalizability.

Conclusions:

• The proposed HEHP network effectively leverages disentangled representations from multimodal data for enhanced underwater salient object detection.
• The developed techniques for handling modal discrepancies, frequency information, and noisy depth data contribute to significant performance improvements and broader applicability.