In this paper, a sparsity-driven auto-focus technique is developed for radar imaging. In order to obtain focused radar imagery, the sparsity of target scene is exploited in a statistical manner. In particularly, scattering coefficients are modeled as unknown random variables in a Bayesian way to encode sparsity and phase errors are modeled as unknown deterministic variables. Moreover, smoothness prior of the phase errors is also imposed in a non-parametric way to regularize the solution space for more accurate estimation. Based on formulated the graphical model, the scattering coefficient and phase errors can be jointly estimated via a variational Bayesian expectation maximization approach, where sparsity of the target scene can be obtained subsequently. The experimental results have validate the effectiveness of the proposed method in terms of imaging quality of the target scene and estimation accuracy of the phase error.