Underwater optical image enhancement via an unsupervised contrastive learning-guided diffusion model

The wavelength-dependent attenuation and absorption of light in water often result in images suffering from color distortion, blurriness, and low contrast. Moreover, the complex seabed terrain and intricate ocean currents make it impractical to capture paired underwater images. To tackle these challenges, an unsupervised contrastive learning-guided diffusion model (UCL-Diff) is proposed for real-world underwater optical image enhancement. This model uniquely integrates adversarial training with a Denoising Diffusion Probabilistic Model (DDPM) for underwater optical image enhancement. Specifically, an Adversarial Contrastive Learning-based Pre-Enhancement Network (ACL-Net) is first proposed to generate sub-optimal quality underwater images from unpaired datasets. Based on these pre-enhanced results, a Discrete Wavelet Transform-based Conditional Diffusion Model (DWT-CDM) is then designed to process the low-frequency information in the wavelet domain, while a High-Frequency Enhancement Module (HFEM) is created to enrich high-frequency information. Lastly, multiple loss functions are explored to optimize both ACL-Net and DWT-CDM. Extensive experiments demonstrate that the proposed UCL-Diff can effectively mitigate issues such as color bias and low contrast, even when using unpaired underwater datasets. The results achieved are comparable to those obtained through supervised training.