Sonar target representation using two-dimensional Gabor wavelet features.

Area of Science:

• Acoustics
• Signal Processing
• Machine Learning

Background:

• Automated target classification (ATC) from sonar data is crucial for various applications.
• Existing methods often struggle with feature extraction and require large datasets.
• Integrating domain knowledge with machine learning can improve ATC performance.

Purpose of the Study:

• To introduce a new feature extraction technique for sonar magnitude spectra.
• To improve automated target classification accuracy and robustness.
• To demonstrate the effectiveness of model-cognizant feature extraction.

Main Methods:

• Feature extraction via convolution of 2D Gabor wavelet with acoustic color magnitudes.
• Identification and culling of informative features (Gabor stripes).
• Application of supervised machine learning classifiers (SVM, Random Forest, Feed-forward Neural Network).

Main Results:

• The Gabor stripe features are target-specific and invariant to aspect angle.
• A threshold-based culling process effectively removes non-informative features.
• Significant classification performance increases were observed, up to 47% with a Random Forest classifier on PondEx10 data.

Conclusions:

• The proposed feature extraction method effectively captures informative target characteristics from sonar data.
• Model-cognizant feature extraction enhances classification accuracy, enabling high performance even with limited data.
• This approach bridges model-based domain knowledge and machine learning for robust ATC.