Self-supervised polarization image dehazing method via frequency domain generative adversarial networks

Haze significantly hinders the application of autonomous driving, traffic surveillance, and remote sensing. Image dehazing serves as a key technology to enhance the clarity of images captured in hazy conditions. However, the lack of paired annotated training data significantly limits the performance of deep learning-based dehazing methods in real-world scenarios. In this work, we propose a self-supervised polarization image dehazing framework based on frequency domain generative adversarial networks. By incorporating a polarization calculation module into the generator, the Stokes parameters of airlight are accurately estimated, which are used to reconstruct the synthesized hazy image by combining the dehazed image generated via a densely connected encoder-decoder. Furthermore, we optimize the discriminator with frequency domain features extracted by frequency decomposition module and introduce a pseudo airlight coefficient supervision loss to enhance the selfsupervised training. By discriminating between synthetic hazy images and real hazy images, we achieve adversarial training without the need for paired data. Simultaneously, supervised by the atmospheric scattering model, our network can iteratively generate more realistic dehazed images. Extensive experiments conducted on the constructed multi-view polarization datasets demonstrate that our method achieves state-of-the-art performance without requiring real-world ground truth.