A Lightweight Network Based on Multi-Scale Asymmetric Convolutional Neural Networks with Attention Mechanism for Ship-Radiated Noise Classification

Chenhong Yan, Shefeng Yan, Tianyi Yao, Yang Yu, Guang Pan, Lu Liu, Mou Wang, Jisheng BaiJournal of Marine Science and Engineering 2024 · DOI

TL;DR

This paper introduces MA-CNN-A, a novel lightweight deep learning model for classifying ship-radiated underwater noise. By employing multi-scale asymmetric convolutions and an attention mechanism, it achieves high accuracy (98.2%) with significantly fewer parameters and faster inference speeds, making it suitable for deployment on resource-constrained embedded systems.

What Problem Does This Paper Address?

Accurately identifying and classifying ship-radiated underwater noise is crucial for maritime security and monitoring. While deep learning, particularly Convolutional Neural Networks (CNNs), has shown promise in this area by analyzing time-frequency representations of sound, existing models are often computationally expensive. This high computational cost makes them impractical for real-time applications on embedded devices, which are common in underwater acoustics due to power and processing limitations. The challenge, therefore, is to develop a model that is both highly accurate and computationally efficient for ship noise classification.

Methods

The core of the MA-CNN-A model lies in its innovative architecture designed for efficiency and effectiveness:

  • Multi-Scale Asymmetric Convolutional Backbone: Instead of standard square convolutional kernels, the model uses asymmetric kernels (e.g., 1x8 and 8x1). This factorization reduces the number of parameters and computations while allowing the network to capture features across different scales and dimensions (time and frequency) of the Mel spectrograms. The backbone is further enhanced by using multiple branches with different kernel sizes (8, 16, 32, 64) to extract features at various resolutions.
  • Improved ECA Attention Mechanism: To further refine feature selection and reduce redundancy, an enhanced Efficient Channel Attention (ECA) module is incorporated. This module adaptively rebalances feature weights across channels, focusing on the most informative ones without significantly increasing model complexity. The paper highlights that the attention mechanism adds a negligible number of parameters (0.02‰ of the total).
  • Feature Representation: The model utilizes Mel spectrograms as input, which are derived from Short-Time Fourier Transform (STFT) and are known to better represent human auditory perception and capture relevant underwater acoustic features.

Datasets

The study utilizes the DeepShip dataset, a publicly available benchmark for underwater acoustics. This dataset comprises approximately 47 hours of real-world underwater recordings from 265 ships, categorized into four types: cargo, passenger ship, oil tanker, and tug. The authors processed this dataset by resampling audio files to 16 kHz, segmenting them into three-second clips, and then generating Mel spectrograms. The dataset was partitioned into training (70%), validation (20%), and test (10%) sets.

Key Insights and Ideas

  • Asymmetric Convolutions for Efficiency: The paper demonstrates that breaking down square convolutions into asymmetric ones (horizontal and vertical) is a powerful strategy for reducing model parameters and computational load without sacrificing accuracy. This is particularly effective in the time-frequency domain where distinct temporal and spectral features are crucial.
  • Multi-Scale Feature Extraction is Key: Using multiple branches with different kernel sizes in the convolutional backbone allows the model to capture a richer set of features at various scales, leading to better discrimination between ship types.
  • Attention for Focused Learning: The improved ECA attention mechanism effectively prunes less relevant feature channels, allowing the model to concentrate on the most discriminative information, thereby boosting performance and robustness.
  • Lightweight Models for Embedded Systems: The research strongly advocates for and successfully demonstrates the feasibility of developing highly accurate deep learning models that are small and fast enough for deployment on resource-constrained underwater devices.

Key Results

  • High Accuracy: The MA-CNN-A model achieved an impressive 98.2% recognition accuracy on the DeepShip dataset, outperforming several state-of-the-art methods.
  • Parameter Reduction: Compared to a standard CNN with three-scale square convolutions, the MA-CNN-A achieved a 68.1% reduction in parameters while improving accuracy.
  • Computational Efficiency: The model boasts a significantly smaller model size (3.9 MB) and fewer parameters (1.00 M) compared to many other models. It also achieves a fast inference speed of 310.2 fps on GPU and 123.5 fps on CPU.
  • Robustness: The MA-CNN-A demonstrated robust performance against varying Signal-to-Noise Ratios (SNRs), maintaining high accuracy even in low-SNR conditions.
  • Ablation Studies: Experiments confirmed that the improvements in accuracy and efficiency are attributed to the combined effect of the multi-scale asymmetric convolution, the multi-scale block, and the attention mechanism.

Limitations and Future Work

  • Dataset Specificity: While the DeepShip dataset is comprehensive, future work could explore the model's performance on other underwater acoustic datasets to assess its generalizability.
  • Real-World Deployment: Although designed for lightweight deployment, further testing and optimization in actual underwater environments with diverse acoustic conditions and potential hardware constraints would be beneficial.
  • Feature Fusion: The paper mentions that the model fuses features from four branches by addition rather than concatenation. Exploring other fusion strategies might yield further improvements.
  • Exploration of Other Asymmetric Convolutions: While the paper focuses on specific asymmetric kernel shapes, investigating a wider range of asymmetric kernel designs could be a future research direction.

Summary updated

Abstract

Ship-radiated noise classification is critical in ocean acoustics. Recently, the feature extraction method combined with time–frequency spectrograms and convolutional neural networks (CNNs) has effectively described the differences between various underwater targets. However, many existing CNNs are challenging to apply to embedded devices because of their high computational costs. This paper introduces a lightweight network based on multi-scale asymmetric CNNs with an attention mechanism (MA-CNN-A) for ship-radiated noise classification. Specifically, according to the multi-resolution analysis relying on the relationship between multi-scale convolution kernels and feature maps, MA-CNN-A can autonomously extract more fine-grained multi-scale features from the time–frequency domain. Meanwhile, the MA-CNN-A maintains its light weight by employing asymmetric convolutions to balance accuracy and efficiency. The number of parameters introduced by the attention mechanism only accounts for 0.02‰ of the model parameters. Experiments on the DeepShip dataset demonstrate that the MA-CNN-A outperforms some state-of-the-art networks with a recognition accuracy of 98.2% and significantly decreases the parameters. Compared with the CNN based on three-scale square convolutions, our method has a 68.1% reduction in parameters with improved recognition accuracy. The results of ablation explorations prove that the improvements benefit from asymmetric convolution, multi-scale block, and attention mechanism. Additionally, MA-CNN-A shows a robust performance against various interferences.
BibTeX
@article{yan2024lightweight,
  author = {Chenhong Yan and Shefeng Yan and Tianyi Yao and Yang Yu and Guang Pan and Lu Liu and Mou Wang and Jisheng Bai},
  title = {{A Lightweight Network Based on Multi-Scale Asymmetric Convolutional Neural Networks with Attention Mechanism for Ship-Radiated Noise Classification}},
  journal = {Journal of Marine Science and Engineering},
  year = {2024},
  doi = {10.3390/jmse12010130},
}
Figure