Zero-UMSIE: a zero-shot underwater multi-scale image enhancement method based on isomorphic features.

Area of Science:

• Computer Vision
• Image Processing
• Deep Learning

Background:

• Underwater images suffer degradation from light scattering and absorption.
• Scarcity of paired real-world data and limitations of synthetic data hinder deep learning-based restoration.
• Restoring degraded underwater images presents a significant challenge.

Purpose of the Study:

• To propose a zero-shot underwater image enhancement method (Zero-UMSIE).
• To address the limitations of paired data scarcity in deep neural network-based image restoration.
• To improve the quality and generalization ability of underwater image enhancement.

Main Methods:

• Estimating global background light, transmission map, and scene radiance from original underwater images.
• Generating re-degraded images by mixing estimated scene radiance with original images.
• Employing multi-scale and non-reference loss functions for network fine-tuning and generalization.

Main Results:

• The proposed Zero-UMSIE method effectively enhances degraded underwater images.
• The method demonstrates superior performance compared to state-of-the-art techniques on real-world datasets.
• Evaluations show significant improvements in image quality, addressing color bias and uneven illumination.

Conclusions:

• Zero-UMSIE offers a robust solution for underwater image enhancement without requiring paired datasets.
• The method exhibits competitive and applicable performance across diverse underwater conditions.
• The approach effectively overcomes common challenges in underwater image restoration.