Underwater Passive Target Classification based on β Variational Autoencoder and MFCC

Underwater passive target classification is an open set classification problem, where quite often test data of those classes, which were not present during training phase is encountered and it is a challenging task due to the intrinsic complexity of the radiated noise from the target. Conventional classification architectures with spectral processing often fail miserably. Supervised learning methods like deep learning, offers higher success rate but they require enormous amount of data for training and their performance in open set classification is again a challenge. This paper presents an effective method for underwater target classification by the beta variational autoencoder (beta - VAE) model with Mel Frequency Cepstral Coefficients (MFCCs) features. MFCC effectively utilises the non-linear auditory effect of the human ear with different frequencies. beta - VAE, being one of the generative models, is capable of generalizing with less amount of data. Classification experiments on various underwater targets have been performed with the proposed method, and results indicate that the proposed method is effective in underwater passive target classification.