Side-Scan Sonar Image Segmentation Based on Multi-Channel CNN for AUV Navigation.

Area of Science:

• Robotics
• Computer Vision
• Marine Technology

Background:

• Autonomous Underwater Vehicle (AUV) navigation relies on sensor data.
• Side-scan sonar provides underwater imagery but its semantic information is underutilized.
• Existing methods lack practical ways to leverage sonar image semantics for AUV navigation.

Purpose of the Study:

• To propose a novel convolutional neural network (CNN) model for extracting semantic information from side-scan sonar images.
• To improve the capabilities of AUV autonomous navigation and mapping.
• To address the lack of practical methods for utilizing sonar image semantics.

Main Methods:

• Developed a CNN model with a standard codec structure.
• Implemented multi-channel feature extraction and fusion to reduce parameters and enhance feature weights.
• Utilized large convolution kernels for large-scale sonar images and added parallel compensation links with small-scale kernels for multi-scale feature extraction.
• Conducted experiments on a self-collected sonar dataset.

Main Results:

• The proposed CNN model achieved high performance with Accuracy (ACC) of 0.87 and Mean Intersection over Union (MIoU) of 0.71.
• The model demonstrated superior results compared to other classical semantic segmentation networks.
• The network is computationally efficient with 347.52 g FOLP and approximately 13 million parameters, ensuring speed and portability.

Conclusions:

• The developed CNN model successfully extracts semantic information from side-scan sonar images.
• This semantic information can significantly assist in AUV autonomous navigation and mapping.
• The model offers an efficient and portable solution for enhancing underwater robotic capabilities.