Optimized UNet framework with a joint loss function for underwater image enhancement

Area of Science:

• Computer Vision
• Image Processing
• Underwater Imaging

Background:

• Underwater images often suffer from low quality due to various environmental factors.
• Enhancing underwater image quality is crucial for applications in water economy, ecology protection, and sustainable development.

Purpose of the Study:

• To address detail blurring, color imbalance, and noise interference in low-quality underwater images.
• To propose an optimized UNet framework with a joint loss function (OUNet-JL) for superior underwater image enhancement.

Main Methods:

• Developed a Multi-Residual Module (MRM) to improve detail feature representation.
• Introduced a Spatial Multi-Scale Feature Extraction Module (SMFM) with channel attention to correct color imbalance.
• Designed a Strengthen-Operate-Subtract (SOSFM) module for noise reduction and distortion correction.
• Integrated four loss functions into a joint loss function for efficient network training.

Main Results:

• The proposed OUNet-JL framework demonstrated superior performance compared to state-of-the-art algorithms on UIEB and UFO-120 datasets.
• Ablation studies confirmed the effectiveness of the individual modules (MRM, SMFM, SOSFM).

Conclusions:

• The OUNet-JL framework effectively enhances underwater image quality by tackling key degradation issues.
• The proposed approach offers a significant advancement in underwater image restoration technology.