Related Experiment Video
Updated: May 28, 2026
03:31
End-To-End Deep Neural Network for Salient Object Detection in Complex Environments
Published on: December 15, 2023
Hybrid Mamba-CNN network for forward-looking sonar image segmentation with acoustic background suppression mechanism
1State Key Laboratory of Submarine Geoscience, Shanghai Jiao Tong University, Shanghai, 200240, China; School of Marine Science and Technology, Northwestern Polytechnical University, Xi'an, 710072, Shaanxi, China.
Summary
MambaSonar, an efficient CNN-Mamba model, enhances Forward-Looking Sonar (FLS) image segmentation. It achieves superior accuracy and computational efficiency for underwater mapping by combining CNNs with Mamba's selective state-space mechanism.
Area of Science:
- Marine technology
- Artificial intelligence
- Image processing
Background:
- Forward-Looking Sonar (FLS) is vital for underwater exploration and mapping.
- Current CNN-Transformer models offer strong FLS segmentation but face computational challenges.
- Self-attention mechanisms in these models lead to high overhead, limiting efficiency.
Purpose of the Study:
- To develop an efficient and accurate model for FLS image segmentation.
- To address the computational limitations of existing CNN-Transformer frameworks.
- To introduce a novel hybrid CNN-Mamba architecture named MambaSonar.
Main Methods:
- Proposed MambaSonar, integrating CNNs for local features and Mamba for global dependencies.
- Introduced an acoustic background suppression block to reduce noise in FLS images.
- Developed a Mamba-CNN fusion block for effective multi-scale feature integration.
Main Results:
- MambaSonar demonstrated superior segmentation accuracy on public FLS datasets.
- The model achieved high computational efficiency compared to existing methods.
- Ablation studies confirmed the effectiveness of MambaSonar's components.
Conclusions:
- MambaSonar offers a promising solution for efficient and accurate FLS image segmentation.
- The hybrid CNN-Mamba architecture shows potential for challenging underwater imaging tasks.
- The proposed acoustic suppression and fusion blocks enhance model performance.