Unsupervised Underwater Image Restoration: From a Homology Perspective

Abstract

Underwater images suffer from degradation due to light scattering and absorption. It remains challenging to restore such degraded images using deep neural networks since real-world paired data is scarcely available while synthetic paired data cannot approximate real-world data perfectly. In this paper, we propose an UnSupervised Underwater Image Restoration method (USUIR) by leveraging the homology property between a raw underwater image and a re-degraded image. Specifically, USUIR first estimates three latent components of the raw underwater image, i.e., the global background light, the transmission map, and the scene radiance (the clean image). Then, a re-degraded image is generated by randomly mixing up the estimated scene radiance and the raw underwater image. We demonstrate that imposing a homology constraint between the raw underwater image and the re-degraded image is equivalent to minimizing the restoration error and hence can be used for the unsupervised restoration. Extensive experiments show that USUIR achieves promising performance in both inference time and restoration quality.

A Lightweight Multi-Branch Context Network for Unsupervised Underwater Image Restoration

“…IBLA [15] 2.4965 15.6838 0.7002 UDCP [8] 6.3967 10.5691 0.5081 ULAP [17] 2.2741 16.0278 0.7484 UWCNN [23] 2.7715 14.7574 0.7023 Water-Net [22] 0.9420 19.5467 0.8339 UGAN [19] 1.0669 18.6064 0.6772 FUnIE-GAN [24] 1.2969 18.8296 0.7542 Ucolor [25] 0.6210 21.2579 0.8460 CWR [28] 0.7650 20.3273 0.8270 USUIR [29] 0.6953 20.5172 0.8373 Peng et al [32] 0 1 shows that the proposed approach gets the highest SSIM on T-U190 and lower MSE and PSNR values compared to Ucolor [25] and CWR [28]. However, it's worth noting that our model has a significantly lower parameter count of only 0.12 M, which is 0.08% of Ucolor [25] and 0.81% of CWR [28].…”

“…We conduct a quantitative evaluation as well as a qualitative comparison between our method and other 11 different methods on several datasets. These 11 methods include 3 traditional methods (IBLA [15], UDCP [8], and ULAP [17]), 7 deep learning-based methods (UWCNN [23], Water-Net [22], UGAN [19], FUnIE-GAN [24], Ucolor [25], USUIR [29], and Peng et al [32]), and 1 contrastive learning-based method (CWR [28]). This comparison helps to evaluate the effectiveness of the proposed method in addressing various challenges associated with underwater image restoration.…”

“…Existing underwater image restoration methods mainly include traditional restoration methods [6][7][8][9][10][11][12][13][14][15][16][17][18] and learning-based restoration approaches [19][20][21][22][23][24][25][26][27][28][29][30][31][32]. Most traditional restoration methods rely on underwater imaging models, utilizing prior knowledge and optical characteristics to estimate unknown parameters and then reconstruct the real underwater scene.…”

“…Learning-based restoration methods explore the relationship between underwater images and unknown model parameters in a synthetic or real-world data-driven manner to obtain more accurate and robust estimates, resulting in more realistic and precise restoration images. Deep learning-based restoration approaches are broadly partitioned into two types: those based on Convolutional Neural Networks (CNN), such as [21][22][23]25,27,[29][30][31], and those based on Generative Adversarial Networks (GAN), such as [19,20,24,26]. In addition, there are a few studies based on contrastive learning, such as [28], and a few research based on visual transformers, such as [32].…”

Synergistic Multiscale Detail Refinement via Intrinsic Supervision for Underwater Image Enhancement

“…(Zhou et al 2023d) uses an augmented U-Net to fuse inputs, such as the original, color-corrected, and contrast-enhanced images, to effectively leverage features. Unsupervised underwater image restoration method (USUIR) (Fu et al 2022) involves designing a transmission subnet, a global background subnet, and a scene radiance subnet to estimate parameters of UIFM, facilitating Photo-realistic image restoration. Zhang et al (Zhang et al 2023) proposed Rex-Net, which applied the Retinex theory to enhance underwater images.…”