A Novel Lightweight Model for Underwater Image Enhancement

Area of Science:

• Computer Vision
• Image Processing
• Deep Learning

Background:

• Underwater images inherently suffer from low contrast and color distortion.
• Existing methods for underwater image enhancement often require substantial storage and computational resources.

Purpose of the Study:

• To propose a lightweight model, Rep-UWnet, for enhancing underwater images.
• To reduce storage and computational requirements for underwater image enhancement.
• To improve the objective and subjective quality of underwater images.

Main Methods:

• A lightweight convolutional neural network architecture (Rep-UWnet) is designed.
• The model incorporates a SimSPPF module for feature extraction and a multi-scale hybrid convolutional attention module for feature reweighting.
• Three densely connected RepConv blocks with skip connections are utilized for efficient feature extraction.

Main Results:

• Rep-UWnet achieves an 83% reduction in parameters (from 2.7M to 0.45M) compared to previous methods.
• The model outperforms state-of-the-art algorithms in objective quality metrics.
• Enhanced image quality is demonstrated through subjective assessments of contrast, colorimetry, and clarity.

Conclusions:

• Rep-UWnet offers a highly efficient solution for underwater image enhancement.
• The proposed model significantly improves visual quality and is effective for downstream high-level vision tasks.
• This lightweight approach provides a practical solution for resource-constrained underwater imaging applications.