Underwater signal recognition based on integrating domain adaptation framework with the stochastic classifier

Although deep learning has made impressive progress in underwater target recognition, most current methods ignore the dataset mismatch caused by various marine conditions. To mitigate the mismatching, this paper designs a domain adaptation framework. The scheme leverages minimax entropy (MME) and conditional adversarial domain adaptation (CDAN) theories to encourage the model to extract domain invariant features. An additional stochastic classifier is introduced, and the consistency of prediction results between classifiers is exploited to determine pseudo-labels. Meanwhile, weights are constructed for poorly aligned data to strengthen the adversarial training. To fully utilize sample information, we contrive the input reconstruction, and combine constant-Q transform (CQT), double delta-constant-Q cepstral coefficient (CQCC) to achieve target recognition. Furthermore, the convolutional neural network (CNN) with the second-order pooling layer (SOP) is modified to ensure the discriminability and transferability of deep features. The results in experiments show that the proposed strategies are beneficial to improve frameworks’ performance. •We develop a network input and apply it with the constant-Q transform-type features.•The effects of models deploying instance normalization are verified.•The scheme integrates two domain adaptation mechanisms for model training.•The stochastic classifier is used to help evaluate the sample prediction.