Canonical correlation analysis as a feature extraction method to classify active sonar targets with shallow neural...

Area of Science:

• Signal Processing
• Machine Learning
• Acoustics

Background:

• Sonar target recognition is complex due to acoustic scatterers, clutter, and propagation effects.
• Variations in target echoes and platform positions pose significant challenges.
• Effective feature extraction is crucial for accurate sonar classification.

Purpose of the Study:

• To investigate sonar target classification using a statistical approach.
• To extract salient target feature vectors for improved classification.
• To apply canonical correlation analysis (CCA) for feature extraction in active sonar data.

Main Methods:

• Employed multivariate statistical method: canonical correlation analysis (CCA).
• Applied CCA with a sliding window to extract maximally correlated projections.
• Used projection feature vectors to train a neural network classifier.
• Utilized confusion matrices and layer-wise relevance propagation for analysis.

Main Results:

• Canonical correlation analysis (CCA) feature vectors increased classification accuracy by 10% to 34% compared to unprocessed features.
• Demonstrated the effectiveness of CCA in capturing persistent, morphing target features.
• Neural network classifier performance was significantly enhanced by CCA-derived features.

Conclusions:

• Canonical correlation analysis (CCA) is a powerful technique for sonar target feature extraction.
• The proposed method offers a statistically robust approach to enhance sonar classification accuracy.
• This statistical feature extraction significantly improves performance in complex acoustic environments.