A hierarchical probabilistic underwater image enhancement model with reinforcement tuning

Abstract

Underwater Image Enhancement (UIE) is a challenging problem due to the complex underwater environment. Traditional UIE methods can hardly adapt to various underwater environments. Deep learning-based UIE methods are more powerful but often rely on a large deal of real-world underwater images with distortion-free reference images. This gives rise to two issues: First, the reference images are highly uncertain because the ground-truth images cannot be are captured directly in underwater environment. Second, learning-based methods may lack generalization ability for diverse underwater environments. To tackle these issues, we propose HPUIE-RL, a hierarchical probabilistic UIE model facilitated by reinforcement learning. This model integrates UNet with hierarchical probabilistic modules to produce various enhanced candidate images that reflect the uncertainty of the enhancement. Then, a reinforcement learning fine-tuning framework is designed to fine-tune the pretrained model in an unsupervised manner, which responds to the dynamic underwater environment. Experiments on real-world datasets from diverse underwater environments demonstrate that our HPUIE-RL model outperforms state-of-the-art UIE methods regarding visual and quantitative performance and generalizability.