Abstract: To address common issues in underwater images, such as color distortion and reduced contrast, as well as the limitations of supervised methods that rely on large-scale paired high-quality underwater image datasets, an unsupervised underwater image enhancement method is proposed. This method utilizes a conditional variational autoencoder (cVAE) combined with probabilistic adaptive instance normalization (PAdaIN) and multi-color space stretching techniques to improve the visual quality of generated images while ensuring consistency with the original input images. Furthermore, a multi-scale residual connection module is employed to effectively reduce the transmission of non-essential information, thereby enhancing the model's performance. This approach provides an alternative to traditional methods that rely on reference images for training.
Experimental results demonstrate that this method achieves a 12% and 3% improvement in Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity Index (SSIM) on the test set compared to FunieGAN and Water-Net, respectively, significantly enhancing the visual quality of the enhanced images. Moreover, the method exhibits excellent performance across different test sets, demonstrating its robust generalization capability. The study indicates that, without the need for reference images, this approach significantly improves underwater image quality, effectively enhancing image detail and color correction, and provides a viable solution for applications in aquaculture and marine monitoring.

Key words: underwater image enhancement, unsupervised learning, multi-scale residual connections, image processing, probabilistic models

摘要： 为解决水下图像常见的颜色失真、对比度降低以及有监督方法在缺乏大规模成对的高质量水下图像数据集支持时效果一般等问题，提出了一种无监督水下图像增强方法，该方法利用条件变分自动编码器（cVAE）结合概率自适应实例归一化（PAdaIN）以及多色空间拉伸技术，旨在提高生成图像的视觉质量，确保生成图像与原始输入图像在视觉上具有一致性。此外，多尺度残差连接模块有效减少了非关键信息的传递，进一步提升了模型的性能。该方法提供了一个以依赖参考图像作为训练数据的替代方案。结果显示，该方法在测试集上的峰值信噪比（PSNR）和结构相似性指数（SSIM）分别比FunieGAN和Water-Net提升12%和3%，显著改善了增强后图像的视觉效果，同时，该方法在不同测试集上的优异表现也验证了其良好的泛化能力。研究表明，该方法在无需参考图像的情况下，显著改善了水下图像的质量，有效提升了图像的细节和色彩校正，为水产养殖和海洋监测提供了一个有效的解决方案。

关键词: 水下图像增强, 无监督学习, 多尺度残差连接, 图像处理, 概率模型