Performance comparison of two types of auditory perceptual features in robust underwater target classification

Robust underwater target classification is an important research topic. Motivated by human auditory principles, the application of auditory perceptual features has attracted increasing attentions. This study compared the robustness of two types of features: one type was directly related to basic auditory attributes, and the other was the central representation obtained by an auditory model. The dataset included nine target types, and evaluation experiments were conducted under conditions corrupted by four types of noises. Then, a support vector machine (SVM) and a deep neural network (DNN) were compared to further investigate the effects of the classifier on feature robustness. The results showed that, while both types of features achieved good performances under noise-free conditions, the central auditory representation was more robust as noise levels increased. Although better results were obtained using the SVM under good conditions, the DNN was less sensitive to noise interference. Including noise-corrupted sounds in the training set further decreased the model sensitivity to noises. Physical analysis indicated that the superiority of the central auditory representation depended on the noise inhibition of peripheral stage and the target/noise separability of central stage. Finally, training with noise-corrupted sounds made the DNN learn compact features and thus improved the robustness.