Degradation-Aware Multi-Stage Fusion for Underwater Image Enhancement.

Area of Science:

• Computer Vision
• Image Processing
• Artificial Intelligence

Background:

• Underwater images suffer from significant degradation including color casts, low illumination, and blur.
• These issues stem from wavelength-dependent absorption and scattering of light underwater.
• Existing enhancement methods may lack efficiency or adaptability for real-time applications on resource-constrained devices.

Purpose of the Study:

• To develop a practical, real-time, two-stage framework for enhancing underwater images.
• To prioritize deployability on edge devices for practical underwater imaging applications.
• To create a degradation-aware system that can handle multiple types of image corruption.

Main Methods:

• A lightweight Convolutional Neural Network (CNN) was used in Stage I for classifying image degradation into three classes (color cast, low light, blur).
• Stage II employed scene-specific enhancement pipelines and output fusion using Linear Fusion and LiteUNetFusion modules.
• LiteUNetFusion utilized spatially adaptive weighting with optional residual correction for refined enhancement.

Main Results:

• The CNN classifier achieved 91.85% accuracy on a subset of the EUVP dataset.
• Linear Fusion improved performance metrics (PSNR, UCIQE, UIQM) significantly compared to single-scene optimizers, with low latency.
• LiteUNetFusion further enhanced results, improving PSNR by +1.5 dB over Linear Fusion and preserving local details in mixed-degradation scenes.

Conclusions:

• The proposed two-stage framework offers effective and efficient real-time enhancement for degraded underwater images.
• The modular design and lightweight components make it suitable for deployment on edge devices.
• Further research is needed to address limitations such as occasional blur/noise amplification and domain generalization.