DiffWater: Underwater Image Enhancement Based on Conditional Denoising Diffusion Probabilistic Model

DiffWater: Underwater Image Enhancement Based on Conditional Denoising Diffusion Probabilistic Model

Underwater imagery suffers from color distortion, low contrast, and haze caused by light absorption and scattering — degrading applications in marine biology, underwater robotics, and offshore inspection. This project, DiffWater, leverages state-of-the-art generative AI and diffusion models to restore underwater images to near-natural quality. The system implements a conditional denoising diffusion probabilistic model (DDPM) that progressively transforms degraded underwater inputs into clear, color-balanced outputs, conditioned on depth-estimation and color-attention maps for physically plausible restoration. Compared to traditional GAN-based and classical filtering approaches, the diffusion framework delivers superior structural fidelity, fewer artifacts, and better generalization to unseen water conditions. Performance is evaluated using PSNR, SSIM, and UIQM metrics on standard underwater benchmark datasets. Implemented in Python with PyTorch and deployed through a web-based upload-and-enhance interface, the project supports batch processing for video frames. This is a premier deep learning computer vision project for students and researchers working on image enhancement using deep learning, generative AI projects, denoising diffusion models, and marine robotics vision systems.

Components

Python 3.8+
PyTorch
Conditional DDPM architecture
OpenCV
Depth-estimation & color-attention modules
Streamlit upload-and-enhance interface

Key Features

• Conditional denoising diffusion restoration of color, contrast and haze
• Depth-estimation and color-attention conditioning for physically plausible results
• Outperforms GAN-based and classical filtering in structural fidelity with fewer artifacts
• Evaluated with PSNR, SSIM and UIQM on standard underwater benchmarks
• Batch processing support for video frames through a web interface

Applications

Marine robotics teams, underwater inspection services, offshore industries and deep learning researchers — a trending topic with strong publication value.