| Summary: | In the realms of the ocean, it becomes a formidable task to detect and classify the passive acoustic targets from the convoluted acoustic mixture confronted by the sonar frontend. Though the advances in deep learning driven by enormity of data and computational infrastructure have resulted in a tremendous leap in performance across various domains, passive sonar target recognition still remains an elusive task for the acoustic as well as signal processing communities. Various channel related artifacts together with the inherent difficulty in obtaining annotated data limit the target records required for training these massive supervised networks so as to yield an optimal performance. This demands models that can generalize well beyond the often sparse training instances. In order to address this issue, generative frameworks can be utilized to model the causal attributes of the target signature so that the network becomes tolerant to the distortions induced by the ambient noise and channel artifacts. This paper exploits the generative modelling capability of an Auxiliary Classifier Generative Adversarial Network (ACGAN) to construct a data-efficient underwater target classifier. These class-conditioned frameworks based on unsupervised representation learning can model the true data distribution using the latent attributes of the training data. In order to make the causal factors of variation more explicit, the raw time domain samples are transformed into joint time–frequency representations using filterbanks initialized at different perceptual scales. Experimental evaluation of the proposed system on target instances collected from diverse locations of the Indian Ocean yields promising results in terms of data efficiency, class confidence and classification accuracy.
|
|---|