Uncertainty Inspired Underwater Image Enhancement (2207.09689v1)

Abstract: A main challenge faced in the deep learning-based Underwater Image Enhancement (UIE) is that the ground truth high-quality image is unavailable. Most of the existing methods first generate approximate reference maps and then train an enhancement network with certainty. This kind of method fails to handle the ambiguity of the reference map. In this paper, we resolve UIE into distribution estimation and consensus process. We present a novel probabilistic network to learn the enhancement distribution of degraded underwater images. Specifically, we combine conditional variational autoencoder with adaptive instance normalization to construct the enhancement distribution. After that, we adopt a consensus process to predict a deterministic result based on a set of samples from the distribution. By learning the enhancement distribution, our method can cope with the bias introduced in the reference map labeling to some extent. Additionally, the consensus process is useful to capture a robust and stable result. We examined the proposed method on two widely used real-world underwater image enhancement datasets. Experimental results demonstrate that our approach enables sampling possible enhancement predictions. Meanwhile, the consensus estimate yields competitive performance compared with state-of-the-art UIE methods. Code available at https://github.com/zhenqifu/PUIE-Net.

Uncertainty Inspired Underwater Image Enhancement: A Summary for Researchers

The paper "Uncertainty Inspired Underwater Image Enhancement" introduces a novel probabilistic approach to underwater image enhancement (UIE), a field challenged by non-reversible degradation of underwater signals and the inherent difficulty in obtaining ground truth high-quality images. Unlike conventional point estimation methods that often suffer from ambiguous reference maps, this paper proposes a distribution estimation followed by a consensus process to address the uncertainty in labeling.

Probabilistic Network Design

The primary contribution is the introduction of PUIE-Net, the first probabilistic network for UIE. PUIE-Net utilizes a combination of Conditional Variational Autoencoders (CVAE) with adaptive instance normalization (AdaIN) to construct an enhancement distribution for degraded underwater images. This modeling approach allows the network to cope with bias introduced in reference map labeling to some extent. By learning the enhancement distribution, diverse potential solutions can be generated, enabling more robust and stable results.

The network architecture of PUIE-Net is based on a modified U-Net, a widely recognized framework for image-to-image tasks, integrated with a new module termed probabilistic adaptive instance normalization (PAdaIN). PAdaIN is devised to capture the appearance distribution of enhancement statistics in the latent space rather than directly adhering to a deterministic mapping.

Strong Numerical Results

The empirical evaluation on two real-world underwater image datasets, including a newly built ambiguous label dataset, demonstrates competitive quantitative performance of PUIE-Net. Specifically, the paper demonstrates superior scores on metrics like SSIM, PSNR, DeltaE, and NIQE compared to state-of-the-art deterministic methods. Furthermore, subjective evaluations using the Mean Opinion Score (MOS) indicate improved perceived quality, reinforcing the validity of the proposed approach.

Implications and Future Work

This paper's contributions lie not only in the novelty of addressing UIE as a distribution estimation problem but also in its theoretical implications for the broader field of low-level vision tasks. By viewing UIE through the lens of probabilistic modeling, it opens pathways for addressing similar enhancement tasks where ground truth is elusive, such as low-light image enhancement, dehazing, and denoising.

Speculatively, the approach could guide future developments in AI where the uncertainty of data labeling poses significant challenges. The consensus process outlined could be adapted for diverse applications where robust estimations from multiple interpretations are preferable over deterministic predictions.

Conclusion

In summary, the method provided in the paper exemplifies a significant stride towards accommodating ambiguity inherent to UIE by utilizing probabilistic modeling coupled with a consensus process. Researchers in the field could leverage the insights and techniques developed here to navigate uncertainties in enhancing visual data, thereby bolstering the robustness of AI models in environments where true data labels are irretrievable.