Compression of a Deep Competitive Network Based on Mutual Information for Underwater Acoustic Targets Recognition

Area of Science:

• Signal Processing
• Machine Learning
• Acoustics

Background:

• Deep neural networks enhance underwater acoustic target recognition with unlabeled data.
• Redundant features in deep learning hinder recognition accuracy and efficiency.

Purpose of the Study:

• To propose a compressed deep competitive network for improved feature extraction from ship radiated noise.
• To address the limitations of redundant features in deep learning for acoustic target recognition.

Main Methods:

• Integrating competitive learning into restricted Boltzmann machine learning for shared weights.
• Employing mutual information-based network pruning to remove redundant parameters and compress the network.

Main Results:

• The compressed deep competitive network achieved 89.1% classification accuracy.
• This represents a 5.3% increase over the deep competitive network and a 13.1% increase over state-of-the-art methods.
• The network successfully extracts fewer, more informative features, enhancing recognition.

Conclusions:

• The compressed deep competitive network effectively improves underwater acoustic target recognition accuracy and efficiency.
• This approach offers a robust solution for feature extraction in limited, noisy acoustic data.
• The method demonstrates superior performance compared to existing techniques.