Variational inference of reverberation suppression exploiting hierarchical Dirichlet process in underwater moving target detection

The moving target detection in active sonar measurement is essential for underwater surveillance. Complex reverberation in a shallow-water environment frequently gives rise to false detection and severely degrades system performance. Low-rank and sparse decomposition, exploiting the low-rank characteristic of reverberation and the sparse nature of moving target across multi-frame range-bearing images, is currently the mainstream method for reverberation suppression. The performance of the reverberation suppression method based on the optimization framework is affected by the manual selection of the regularization parameter that balances the low-rank reverberation and the sparse moving target. To address the issue, this paper implements reverberation suppression in underwater moving target detection within a Bayesian framework. A hierarchical Dirichlet process with a Gaussian mixture model is developed to characterize the non-low-rank component, which exhibits a non-independent and non-identical distribution across different time frames. The steady reverberation component is modeled using the inherent low-rank structure. Based on the proposed model, variational inference is employed to estimate all involved variables, in which the moving target is effectively extracted from complex reverberation. The superiority and robustness of the proposed method are validated through a field target detection experiment compared with other methods.