Accurate estimation of surface normal plays a vital role in various computer vision tasks. Physically-based shape from polarization methods have limitations restricting their applications. Conversely， learning-based shape from polarization methods outperform physical methods in both accuracy and applicability. To further improve the accuracy of shape from polarization and make it applicable to a broader range of practical tasks， we proposed a novel method. First， we introduced a new polarization information representation combining Stokes vectors， enhancing the model's ability to extract polarization physical prior information. Then， we integrated a bi-level routing sparse self-attention mechanism to improve the model's perception of global contextual information， enabling better disambiguation of local polarization information. Testing on the DeepSfP dataset and out test data， experimental results demonstrate our proposed method achieves an average angular error of 13.37° on the DeepSfP dataset， outperforming existing methods in all tested metrics including accuracy and angular error. This indicates a significant improvement in normal estimation effectiveness with our proposed method. By introducing a novel polarization information representation and sparse self-attention mechanism， our approach enhances the accuracy and applicability of polar surface normal estimation， providing stronger support for practical task applications.