Linear Shrinkage Coefficient-Based Source Number Estimation Using Semi-Supervised GAN With Small Samples

A deep learning-based source number estimation method is presented in this article, where the deep generative adversarial network (GAN) combined with semi-supervised learning is applied, modifying the classifier in an adversarial way. Different from the traditional eigenvalue-based methods, the linear shrinkage coefficient established under the general asymptotic theory framework is utilized as the input feature of the network, which produces more distinct classification features, and therefore achieves satisfactory classification performance under conditions of small number of labels and samples, and low signal-to-noise ratios. It is shown that 30$\%$ label rate is able to achieve a performance close to fully supervised learning.