Unsupervised underwater image restoration via Koschmieder model disentanglement

Most deep learning-based underwater image enhancement approaches rely on paired underwater datasets. Given the scarcity of real-world paired datasets and the inadequacy of synthetic datasets in approximating real-world conditions, using deep learning to restore these degraded underwater images remains challenging. To address this, this paper proposes an unsupervised underwater image restoration method based on the Koschmieder light scattering model, called KM-UUIR. Specifically, we disentangle the input image into the global background light, the scene radiance (the clean image), and the transmission map. These three components are recombined to form a reconstructed image, with a reconstruction consistency constraint applied between the original and reconstructed images. Additionally, the estimated scene radiance and global background light are combined with the original image to generate a re-degraded image. The re-degraded image shares the same formation model and scene radiance as the original image. Thus, a homology consistency constraint is applied between the original and re-degraded images. Furthermore, no-reference constraints are introduced to reduce color distortion and achieve higher quality underwater images. Quantitative and qualitative experiments demonstrate the effectiveness of the proposed method, achieving a PSNR of 24.64 and an SSIM of 0.88 on the UIEB dataset, showing competitive performance even compared to supervised learning methods.