Meta PID Attention Network for Flexible and Efficient Real-World Noisy Image Denoising

Area of Science:

• Computer Vision
• Artificial Intelligence
• Signal Processing

Background:

• Deep convolutional neural networks (CNNs) excel at image denoising using external training data.
• Supervised methods struggle with generalization when test data differs from training conditions, leading to overfitting and performance degradation.

Purpose of the Study:

• To propose a novel denoising algorithm, the Meta PID Attention Network (MPA-Net), to overcome generalization limitations.
• To achieve robust image denoising across various noise types and conditions.

Main Methods:

• MPA-Net is constructed by integrating Meta PID Attention Modules (MPAMs).
• Each MPAM employs a second-order attention module (SAM) for channel-wise feature correlation analysis.
• A proportional-integral-derivative (PID) guided meta-learning framework adaptively updates features and dynamically generates filter weights.

Main Results:

• MPA-Net demonstrates superior denoising performance compared to state-of-the-art methods.
• The network exhibits robust generalization capabilities on diverse, unseen noise conditions across ten datasets.
• Qualitative and quantitative experiments validate the effectiveness of the proposed denoising approach.

Conclusions:

• The proposed MPA-Net effectively addresses the generalization challenges in supervised image denoising.
• The PID-guided meta-learning framework enables adaptive feature learning and robust performance on real-world noisy images.