An Underwater Image Denoising Method Based on High-Frequency Abrupt Signal Separation and Hybrid Attention Mechanism

Area of Science:

• Computer Vision
• Image Processing
• Robotics

Background:

• Underwater image quality is degraded by light absorption and scattering, leading to blurring and noise.
• Poor image quality hinders research and applications relying on underwater visual data.
• Effective underwater image denoising is crucial for reliable data acquisition.

Purpose of the Study:

• To introduce HHDNet, a novel method for underwater image denoising.
• To address noise issues specific to underwater robot photography.
• To improve the quality of underwater images for subsequent analysis.

Main Methods:

• HHDNet employs a dual-branch network architecture to process high and low frequency components separately.
• A hybrid attention module within the Global Context Extractor (GCE) targets high-frequency abrupt noise.
• Gaussian kernel decomposition and efficient residual convolutional units are utilized.

Main Results:

• HHDNet demonstrates superior denoising effectiveness compared to existing methods.
• The method successfully removes high-frequency abrupt noise while preserving image details.
• HHDNet maintains computational efficiency, offering flexibility in application.

Conclusions:

• HHDNet provides an effective solution for underwater image denoising.
• The proposed method enhances image quality for underwater robot photography.
• HHDNet offers a flexible and efficient approach to underwater noise removal.