水下目标信号的结构化稀疏特征提取方法

To extract noise-robust feature of underwater target signal for target recognition, the time-frequency spectrum analysis was used, and the high-energy narrow band line spectra which provide distinguishable information in underwater target recognition were found in most cases, and they were sparsely distributed in the time-frequency spectra of the noise emitted from the underwater targets. On the basis of sparse decomposition theory, a sparse feature extraction method is proposed, which combines the structured sparse characteristics. This method adopts sparse Bayesian learning model, which can utilize the correlation information between adjacent frame samples to strengthen the narrow band line spectra and improve the noise robustness of the proposed feature extraction method. An experiment based on a measured dataset was conducted, and its result showed that the proposed feature is robust to noise. Moreover, it achieved a high recognition performance when the test samples and the train samples were in mismatched noise conditions.