Liu Tong, Zhu Kaiyan, Cao Weiye, Shan Bolin, Guo Fangyi
Opt Express. 2024 Nov 4;32(23):40398-40415. doi: 10.1364/OE.538120.
Due to the scattering and absorption of light, underwater images often exhibit degradation. Given the scarcity of paired real-world data and the inability of synthetic paired data to perfectly approximate real-world data, it's a challenge to restore these degraded images using deep neural networks. In this paper, a zero-shot underwater multi-scale image enhancement method (Zero-UMSIE) is proposed, which utilizes the isomorphism between the original underwater image and the re-degraded image. Specifically, Zero-UMSIE first estimates three latent components of the original underwater image: the global background light, the transmission map, and the scene radiance. Then, the estimated scene radiance is randomly mixed with the original underwater image to generate re-degraded images. Finally, a multi-scale loss and a set of tailored non-reference loss functions are employed to fine-tune the underwater image and enhance the generalization ability of the network. These functions implicitly control the learning preferences of the network and effectively address issues such as color bias and uneven illumination in underwater images, without the need for additional datasets. The proposed method is evaluated on three widely used real-world underwater image datasets. Extensive experiments on various benchmarks demonstrate that the proposed method is superior to state-of-the-art methods subjectively and objectively, which is competitive and applicable to diverse underwater conditions.
由于光的散射和吸收，水下图像常常呈现出退化现象。鉴于配对的真实世界数据稀缺，且合成的配对数据无法完美逼近真实世界数据，利用深度神经网络恢复这些退化图像是一项挑战。本文提出了一种零样本水下多尺度图像增强方法（Zero-UMSIE），该方法利用原始水下图像与重新退化图像之间的同构性。具体而言，Zero-UMSIE首先估计原始水下图像的三个潜在成分：全局背景光、传输图和场景辐射度。然后，将估计出的场景辐射度与原始水下图像随机混合以生成重新退化图像。最后，采用多尺度损失和一组定制的无参考损失函数对水下图像进行微调，并增强网络的泛化能力。这些函数隐含地控制网络的学习偏好，并有效解决水下图像中的颜色偏差和光照不均匀等问题，而无需额外的数据集。所提出的方法在三个广泛使用的真实世界水下图像数据集上进行了评估。在各种基准上进行的大量实验表明，所提出的方法在主观和客观上均优于现有方法，具有竞争力且适用于各种水下条件。
