Underwater Acoustic Target Recognition Based on the Fusion of Morphological Component Analysis and Multi-Discrimination Generative Adversarial Network

A B S T R A C TThis work considers the recognition of underwater acoustic targets, focusing on the well-known convolutional neural network recognition algorithm. This recognition method is based on the deep learning method, and the most critical step is to extract characteristic information from the target signal, which is used as the basis for recognition. However, due to the complexity of the underwater environment, multiple types of targets can no longer be distinguished well based on single feature information alone, especially when sample data are scarce. The question we address in this paper is whether multi-feature fusion is better than single features of underwater acoustic target recognition tasks when the sample data is relatively small. We propose an underwater acoustic target recognition method that combines fusion features and sample expansion to answer this question. More specifically, the target features are sparsely represented in this method based on morphological component analysis (MCA) theory. Then the fusion features of the target are constructed on this basis, which can overcome the drawback of the inadequate description of the target by a single feature. At the same time, we propose the G-feature of the target, a method of binarization, which can normalize and simplify the features of the target. In order to reduce the over-fitting problem of convolution recognition networks because of the scarcity of sample data, we use a multi-discriminant generative adversarial network to expand data samples and introduce a soft voting decision method in the discriminator. Finally, we redesigned a simplified convolutional recognition network to reduce delay and facilitate model deployment. The experimental test is carried out on the ship radiated noise data set, and the result proves that the proposed fusion features recognition method has a higher recognition accuracy rate than a single feature