TL;DR
This paper introduces DRACNN, a novel deep learning model that directly processes raw time-domain underwater acoustic signals for target recognition, bypassing traditional time-frequency analysis. DRACNN achieves high accuracy (97.1%) with significantly fewer parameters and computations than existing methods, making it suitable for resource-constrained underwater systems.
What Problem Does This Paper Address?
Underwater Acoustic Target Recognition (UATR) is crucial for both civilian and military applications, but existing methods often struggle with noisy environments, signal distortion, and computational complexity. Many current deep learning approaches rely on time-frequency representations (like spectrograms) of acoustic signals. However, this process can lose valuable information and requires significant computational resources, hindering real-time applications and deployment on smaller devices. The authors aim to overcome these limitations by developing a more efficient and accurate UATR model that operates directly on the raw time-domain signals.
Methods
The core of this work is the Deep Residual Attention Convolutional Neural Network (DRACNN). The model is designed to process raw time-domain acoustic signals directly, eliminating the need for time-frequency transformations. Key components include:
- Residual Attention Convolution Blocks (RACB): These are the building blocks of DRACNN. Each RACB incorporates:
- Convolutional Layers: For initial feature extraction from the time-domain signal.
- Channel Attention Mechanism (Squeeze-Excitation module): This adaptively weights feature maps from different channels, allowing the model to focus on the most discriminative features.
- Residual Connections: These help mitigate the vanishing gradient problem in deep networks, improving training stability and enabling deeper architectures.
- Max Pooling: Used for down-sampling and feature reduction.
- Deep Convolutional Blocks (DCB): These blocks are used to fuse features from adjacent RACBs and perform further down-sampling, preserving essential signal characteristics.
- End-to-End Architecture: The DRACNN model takes the preprocessed time-domain signal as input and directly outputs the target classification, simplifying the pipeline.
Datasets
The authors utilized two publicly available datasets for their experiments:
- ShipsEar Dataset: This dataset was used for training and evaluating the DRACNN model. It contains audio recordings of 11 vessel types and background noise, collected in Vigo Harbor, Spain. The data was categorized into five classes (A-E) based on vessel length. The authors used 80% for training and 20% for testing, noting the dataset's inherent imbalance.
- DeepShip Dataset: This dataset was used to evaluate the generalization ability of the DRACNN model. It comprises recordings of four types of vessels: Cargo, Passenger ship, Tanker, and Tug. A subset of this data was used for testing.
Key Insights and Ideas
- Direct Time-Domain Processing is Viable and Beneficial: The paper demonstrates that processing raw time-domain signals directly, rather than relying on time-frequency representations, can lead to more efficient and accurate UATR. This bypasses information loss inherent in time-frequency transformations.
- Attention Mechanisms Enhance Feature Selectivity: The integration of channel attention (SE module) within the residual blocks allows the model to dynamically focus on the most informative parts of the acoustic signal, improving its ability to distinguish targets even in noisy conditions.
- Residual Networks Improve Stability and Depth: The use of residual connections is crucial for training deep convolutional networks, preventing performance degradation and enabling the model to learn more complex hierarchical features.
- Efficiency is Key for Practical Deployment: The DRACNN model achieves state-of-the-art accuracy with a remarkably low number of parameters and floating-point operations, making it highly suitable for deployment on resource-constrained platforms like micro-UATR systems or sonar devices.
Key Results
- High Accuracy on ShipsEar: The DRACNN model achieved an average recognition accuracy of 97.1% on the ShipsEar dataset, outperforming ResNet-18 and being comparable to more complex models like AResNet and UTAR-Transformer.
- Significant Computational Efficiency: DRACNN boasts a significantly reduced model size and computational cost. It has approximately 1/36th the parameters of AResNet and 1/10th of UTAR-Transformer, with 1/292nd and 1/46th the floating-point operations, respectively.
- Robustness to Noise: The model's accuracy remained high even under varying signal-to-noise ratios (SNRs), demonstrating its resilience to environmental noise. Accuracy exceeded 90% for SNRs greater than 0 dB.
- Good Generalization: When pre-trained on ShipsEar and tested on the DeepShip dataset, the DRACNN model achieved 89.2% accuracy, indicating its ability to generalize to unseen data and different acoustic environments.
Limitations and Future Work
- Dataset Imbalance: The authors acknowledge that the ShipsEar dataset is imbalanced, with some categories having significantly fewer samples than others. This can affect training and potentially lead to suboptimal performance for underrepresented classes.
- Future Work Directions:
- Data Augmentation: Exploring simulation data or data augmentation techniques to expand datasets and improve performance, especially for imbalanced samples.
- Auditory Attention Mechanisms: Investigating auditory attention-inspired mechanisms to enhance model interpretability and recognition performance in complex target working conditions.
Summary updated