Path-Routing Convolution and Scalable Lightweight Networks for Robust Underwater Acoustic Target Recognition.

Area of Science:

• Marine acoustics
• Underwater acoustics
• Signal processing

Background:

• Accurate vessel identification is crucial for maritime surveillance and ocean protection.
• Current deep learning models for underwater acoustic recognition are computationally intensive and struggle with multi-scale features, limiting their use on resource-constrained devices.

Purpose of the Study:

• To develop an efficient deep learning model for accurate underwater vessel type identification.
• To address limitations of existing models regarding parameter count and multi-scale feature extraction.

Main Methods:

• Proposed a novel path-routing convolution mechanism with multi-dilation-rate parallel paths and an adaptive routing strategy.
• Designed the MobilePR-ConvNet architecture with systematic width scaling for hardware adaptability.
• Conducted experiments on the DeepShip and ShipsEar datasets.

Main Results:

• Achieved high recognition accuracies of 98.58% on DeepShip and 97.82% on ShipsEar.
• Demonstrated robust performance with 77.8% accuracy under low signal-to-noise ratio (10 dB) conditions.
• Validated cross-dataset generalization capabilities in complex marine environments.

Conclusions:

• The proposed MobilePR-ConvNet offers an effective solution for intelligent vessel identification on resource-constrained underwater devices.
• The novel path-routing convolution mechanism enables discriminative extraction of cross-scale acoustic features.
• The model shows strong performance and adaptability for practical maritime surveillance applications.