Haze reduces scene brightness, weakens contrast, and blurs distant objects, thereby degrading visual perception and the reliability of computer vision tasks. Conventional dehazing methods often suffer from overexposure and distortion in bright sky regions, especially in maritime scenes with large sky areas, where sea-surface reflections further increase interference. In addition, complex scenes such as urban and maritime environments usually exhibit uneven illumination and significant depth variations, making stable dehazing difficult. To address these challenges, this study proposes a multi-scene image dehazing method based on polarization principal structure modeling.

The proposed polarization-guided dehazing method mainly consists of three steps: (a) A local entropy map was constructed and the Otsu algorithm was used to adaptively determine thresholds for selecting the low-texture regions, combined with region growing to obtain an initial sky mask. Subsequently, the boundary refinement was performed using gradient-domain polarization-guided filtering, together with morphological operations and small-region removal to construct a structurally stable, high-confidence sky mask, providing reliable prior information for subsequent estimation of atmospheric light at infinity. (b) Local block matrices of the degree of linear polarization are established within the segmented sky region. Bayesian robust principal component analysis (BRPCA) is used for low-rank decomposition to capture the principal sky structure, while a Gaussian mixture model (GMM) is introduced to model the noise term with multiple components so as to distinguish noise or anomalies with different statistical properties. Furthermore, Markov random field (MRF) constraints are incorporated to smooth the Gaussian component assignment matrix and enforce spatial consistency. This combined modeling framework enables stable and accurate estimation of atmospheric light at infinity. (c) An edge-optimized polarization dark channel was introduced to estimate the transmission map. This method effectively reduced the halo effect at object boundaries and maintained structural clarity. Finally, the dehazed image was reconstructed using the atmospheric scattering model.

To comprehensively evaluate the adopted polarization-guided dehazing method, this study conducted experiments on three datasets: urban haze images, maritime scenes with strong reflections, and the Polarlitis benchmark dataset, representing complex structures, reflection interference, and diverse road scenes, respectively. The method was compared with typical non-polarization methods (DCP, Meng’s, Salazar’s, IDE), a classical polarization dehazing method (Schechner’s), and a recent polarization dark channel approach (Wang’s). The quantitative results (Table 1～Table 3) showed that the proposed method achieved the best overall performance across various evaluation indicators. Compared with the original hazy images, information entropy, average gradient, and local standard deviation were significantly increased, while NIQE and BRISQUE were obviously decreased. Specifically, in urban scenes, information entropy and average gradient increased by 8.1% and 1.136 times. In maritime scenes, the increases were 22.1% and 77.3%, respectively. On the Polarlitis dataset, the increases were 21.0% and 3.517 times, respectively, all outperforming the compared methods. Visual results (Fig.5～Fig.7) further demonstrated the advantages of this method. It restored natural brightness and fine edges in urban scenes while avoiding artifacts common in IDE or Schechner’s; effectively suppressed sea-surface reflections and preserved sky–horizon consistency in maritime scenes; and provided clear visibility in overcast hazy conditions on Polarlitis, where Salazar’s and DCP performed poorly. Overall, the method demonstrated good robustness and generality under multi-scene hazy conditions.

This study proposes a multi-scene dehazing method based on polarization principal structure modeling to solve the problems of atmospheric light estimation error and weak generalization in polarization image dehazing. It shows comprehensive advantages in no-reference metrics such as information entropy, average gradient, and NIQE, significantly enhancing details and contrast while suppressing sky overexposure and edge blurring. In addition, the method exhibits good adaptability, strong generalization ability, and robustness across urban haze, sea fog, and other complex environments, indicating promising potential for practical engineering applications. Nevertheless, its performance under nighttime and highly non-uniform illumination conditions still requires further improvement.