Abstract: To address the issue of low estimation performance of the traditional off-grid sparse Bayesian learning algorithm in the complex shallow-water localization environment for acoustic target direction estimation, this paper proposes a real-domain out-of-state sparse Bayesian learning algorithm that combines dictionary learning and unitary transformation for direction estimation. The algorithm employs the K-means singular value decomposition dictionary learning method to represent the actual received signal of a uniform linear array using a small number of linear combinations of basic received signals, thereby achieving noise reduction for the original signal. The denoised signal matrix is then constructed into a processing matrix that satisfies the central Hermitian property. By applying a unitary transformation, the signal data is converted from complex-domain operations to real-domain operations, which reduces computational complexity. Finally, singular value decomposition and outlier sparse Bayesian learning algorithms are used for iterative processing to achieve target direction estimation. Simulation analysis and sea trial data results demonstrate that compared with the off-grid sparse Bayesian learning algorithm, under conditions of low signal-to-noise ratio and low frame rate, the proposed algorithm has improved azimuth estimation accuracy and algorithm robustness, and is less complex.