In Situ Adaptive Feature Extraction for Underwater Target Classification

This research compares the performance improvements of image-based sonar target classification algorithms when they are adapted to changing clutter environments. The distribution of seabed pixels in the sonar imagery is modeled as a correlated, K-distributed random variable allowing for a quantitative representation of seabed environments in the various testing scenarios. Parameterized environments comprising various target-like seabed textures are generated synthetically and used to examine adaptive classification performance. Results demonstrate that optimizing classifier parameters respective to specific environments improves overall classification performance compared to optimizing classifier parameters against a pooled dataset that includes all possible environments.