Underwater polarization image enhancement based on low-rank polarization tensor model

Abstract

Polarization imaging techniques are widely used in underwater image enhancement due to their unique physical properties. Accurate estimation of the degree of polarization (DoP) of backscatter is crucial for the effectiveness of polarization-based underwater image enhancement methods. However, most existing methods often rely on the assumptions that target reflections are unpolarized and that the DoP of backscatter remains spatially constant. These can lead to substantial errors in the DoP of backscatter estimation, adversely affecting the quality of image recovery. To address these limitations, we propose a novel underwater polarization image enhancement method, called LRPT, based on the low-rank polarization tensor model. This method uniquely integrates the influence of target reflection polarization on imaging results. By expressing the DoP of polarized light as a linear combination of the DoP of backscatter and that of the target, LRPT allows for a more accurate estimation of the three-dimensional DoP of the backscatter matrix at the pixel level, rather than assuming a constant backscatter polarization value. Furthermore, to enhance the generalization ability of the method, we designed a dual orthogonal stretch correction module, which effectively resolves color distortion issues in color-polarized images while preserving the polarization relationships among orthogonally polarized images. Numerous experiments demonstrate that the LRPT method not only mitigates problems related to low contrast and color distortion in underwater images but also exhibits robust performance and strong generalization capabilities.