Summary: The field of blind image deblurring has been broadly investigated and has yielded significant achievements. The two-stage method has been widely adopted for blind image deblurring. In particular, the accurate estimation of the blur kernel during the first stage is critical to overall success. Nevertheless, many existing methods may not accomplish sufficient accuracy in the blur kernel estimation, leading to restored images that exhibit boundary distortions and other undesirable artifacts in the second stage. In this paper, a robust blind image deblurring model is proposed, which addresses the inherent uncertainty in the blur kernel estimation during the blind deblurring, decomposing the blur kernel into a deterministic component and a random component for the purpose of reducing the impact of the kernel error on image restoration through iterative estimation. Furthermore, the utilization of the \(L_p\)-norm in image restoration has demonstrated exceptional performance, therefore, the \(L_p\)-norm is utilized to achieve optimal restored images. The effectiveness of the proposed method is investigated through quantitative and qualitative experimental evaluations, which demonstrate its superior performance compared to state-of-the-art methods in benchmark datasets and text images, as well as natural degradation images.