Featureless classification for active sonar systems

Active sonar systems depend on classification algorithms to identify target echoes and suppress false alarms. Historically, classifiers use a set of empirically derived features that exhibit some statistical separation between background clutter and target echoes. In an ideal scenario, these features would form a sufficient set of statistics capturing all of the information required to classify an echo return. Unfortunately, due to their empirically derived nature features are rarely provably sufficient. To overcome this drawback we present a featureless classifier. Instead of features, we use the raw data samples which form a trivial but provable set of sufficient statistics. To classify the raw data, we use the Adaptive Cosine Estimate algorithm which has a history of success with featureless classification in other applications. We will provide an in-depth look at our featureless algorithm that will include performance results on sea-test data from the Malta Plateau database.