Sonar Target Response Feature Extraction Using Neighbourhood Component Analysis

In the field of autonomous sonar target recognition, researchers have long grappled with the challenge of creating reliable representations of sonar target features. The features crucial for distinguishing sonar targets undergo nonlinear changes over time due to unknown, complicated target scattering physics. It becomes essential for any feature extraction algorithm to mitigate these effects while still understanding why features were chosen for improving discrimination. This work investigates using the supervised dimensionality technique, Neighborhood Component Analysis (NCA), for feature extraction. More specifically, we introduce a novel algorithm that uses a cascade of NCA blocks to extract features from sonar target responses with multiple beams. The effectiveness of feature extraction is evaluated with the popular classifiers
$k$
-nearest neighbors
$k$
NN) and support vector machines (SVM). A subset of the labeled examples from the Malta Plateau database were used to evaluate the proposed algorithm.