Few-shot learning for joint model in underwater acoustic target recognition

Area of Science:

• Marine acoustics
• Artificial intelligence
• Machine learning

Background:

• Underwater acoustic target recognition faces challenges due to limited high-quality labeled data for training deep neural networks.
• Conventional few-shot learning methods are not directly applicable in this domain.

Purpose of the Study:

• To propose a novel learning framework for underwater acoustic target recognition models with few samples.
• To enhance model fine-tuning performance by leveraging unlabeled data through a semi-supervised approach.

Main Methods:

• Developed a self-supervised learning framework tailored for underwater acoustic target recognition.
• Implemented a semi-supervised fine-tuning method that mines and labels unlabeled samples based on deep feature similarity.
• Established performance baselines using small sample datasets with varying labeled data amounts.

Main Results:

• The proposed framework significantly improved the recognition performance of four underwater acoustic target recognition models compared to baselines.
• One joint model demonstrated an accuracy increase of 2.04% to 12.14% over baselines.
• Model performance using only 10% of labeled data surpassed performance on the full dataset, reducing reliance on labeled samples.

Conclusions:

• The developed framework effectively alleviates the problem of insufficient labeled samples in underwater acoustic target recognition.
• The approach enhances model robustness and accuracy in data-scarce underwater acoustic environments.