In order to solve the problems of color shift and incomplete dehazing after image dehazing, this paper proposes an improved image self-supervised learning dehazing algorithm that combines polarization characteristics and deep learning. First, based on the YOLY network framework, a multiscale module and an attention mechanism module are introduced into the transmission feature estimation network. This enables the extraction of feature information at different scales and allocation of weights, and effectively improves the accuracy of transmission map estimation. Second, a brightness consistency loss based on the YCbCr color space and a color consistency loss are proposed to constrain the brightness and color consistency of the dehazing results, resolving the problems of darkened brightness and color shifts in dehazed images. Finally, the network is trained to dehaze polarized images based on the atmospheric scattering model and loss function constraints. Experiments are conducted on synthetic and real-world data, and comparisons are made with six contrasting dehazing algorithms. The results demonstrate that, compared to the contrastive dehazing algorithms, the proposed algorithm achieves PSNR and SSIM values of 23.92 and 0.94, respectively, on synthetic image samples. For real-world image samples, color restoration is more authentic, contrast is higher, and detailed information is richer. Both subjective and objective evaluations show significant improvements. This validates the effectiveness and superiority of the proposed dehazing algorithm.