基于对比学习的无监督水下图像快速增强

Rapid unsupervised underwater image enhancement based on contrastive learning

摘要:

目前，基于合成数据集训练的有监督增强模型难以精确模拟水下成像的物理机制，难以适应复杂多变的实际场景；无监督模型多依赖循环一致性框架，结构复杂且效率较低。为此，本文提出了一种基于对比学习的无监督快速水下图像增强算法，通过从浑浊域到清晰域的单向风格转换实现高效增强。算法采用沙漏型瓶颈结构、多特征选择和亚像素融合构建生成器，在保障特征提取性能的同时降低复杂度并加速推理速度；优化的对比损失正负样本提取方法提升了模型对多变水下数据的适应性。实验表明，算法的PSNR和UIQM分别提升6.5%和5.3%，推理时间缩短至0.04秒，GPU资源占用率减少至少65%，在性能与复杂度间实现了良好平衡。

Abstract:

Supervised enhancement models trained on synthetic datasets struggle to accurately simulate the physical mechanisms of underwater imaging, limiting their adaptability to complex real-world scenarios. Unsupervised models, often based on cycle consistency frameworks, tend to suffer from excessive structural complexity and low efficiency. To address these issues, this paper proposes an unsupervised rapid underwater image enhancement algorithm based on contrastive learning, achieving efficient enhancement through unidirectional style transfer from the turbid domain to the clear domain. The algorithm employs a hourglass bottleneck structure with multi-feature selection and sub-pixel fusion to build the generator, ensuring high feature extraction performance while reducing complexity to accelerate inference speed. An optimized contrastive loss with improved positive and negative sample extraction enhances the model’s adaptability to diverse underwater datasets. Experimental results demonstrate that the proposed algorithm improves PSNR and UIQM by 6.5% and 5.3%, respectively, reduces inference time to 0.04 seconds, and decreases GPU usage by at least 65%, achieving a strong balance between performance and complexity.