Objective Existing infrared polarization image enhancement methods have made some progress in cloud suppression, background removal, and image quality improvement. However, there's still room for further research in restoring infrared polarization images under complex cloud backgrounds. On one hand, some methods mainly rely on features like Stokes parameter combinations, degree of polarization, or polarization angle, with relatively limited consideration of the physical differences between cloud scattering background and target radiation. As a result, in cases with complex cloud textures, strong background non-uniformity, or low local contrast, there can still be issues with cloud residue or insufficient detail recovery. On the other hand, as infrared polarization imaging systems gradually expand to mobile platforms, embedded devices, and portable terminals, algorithms need to balance cloud removal effectiveness with computational complexity, storage overhead, and real-time processing capability. The focus here is on the problem that detecting small infrared targets under complex sky backgrounds is easily interfered with by cloud clutter, leading to high false alarm rates.

Methods First, the concept of dark channel prior is introduced, and a pseudo-dark channel statistic suitable for infrared polarization single-channel images is constructed. A pseudo-dark channel statistic based on the local minima of a single channel is developed to estimate the cloud scattering component in infrared sky scenes, providing a basis for subsequent atmospheric scattering restoration. Then, a framework for infrared polarization cloud removal combining polarization decomposition and atmospheric scattering models is established. To address the issue of coupling between cloud scattering components and effective radiative information in infrared polarization images under complex cloud backgrounds, polarization imaging characteristics are integrated with an atmospheric scattering degradation model. Polarization decomposition is employed to separate and enhance polarization-related information in the images, obtaining an initial representation of effective information; the atmospheric scattering model is then used to estimate and suppress the residual cloud background. Finally, a guided filtering optimization method introducing a planarity constraint is designed. For complex cloud edge regions where transmission estimation is easily affected by local texture variations and sudden luminance changes, a planarity constraint parameter is incorporated into the regularization term of guided filtering to adaptively adjust the smoothing degree in different regions. This method helps improve the spatial consistency of transmission estimation in locally flat areas and reduces over-smoothing in areas with structural changes, thereby enhancing the balance between cloud suppression and image detail preservation to a certain extent.

Results and Discussions In complex scenarios with overlapping cloud layers, the contrast (C) and local signal-to-noise ratio (LSCR) of the proposed algorithm are significantly superior to mainstream algorithms such as PFE and DPI. Meanwhile, the algorithm's single-frame runtime is only 0.036 s, providing high-quality input for subsequent spatial-domain target detection tasks.

Conclusions To address the problem of infrared small targets being easily disturbed by cloud clutter in complex sky backgrounds, an infrared polarization cloud removal algorithm based on an atmospheric scattering physical model is proposed. This method combines polarization decomposition with atmospheric scattering modeling, utilizes a pseudo-dark channel prior to estimate the cloud scattering component, and optimizes the transmittance through improved guided filtering, thereby suppressing cloud clutter while preserving target edge information. Experimental results show that the proposed method demonstrates good target enhancement and background suppression capabilities under complex cloud backgrounds. Its contrast and local signal-to-noise ratio outperform methods such as PFE and DPI, while also exhibiting favorable real-time performance. This method can provide high-quality inputs for subsequent infrared small target detection. Future research will further investigate its adaptability under complex weather conditions such as fog and rain, and deployment verification will be conducted on embedded platforms.