Physics-Informed Generative Adversarial Networks for Laser Speckle Noise Suppression.

Area of Science:

• Optics and Photonics
• Image Processing
• Computational Imaging

Background:

• High-resolution microscopy often uses ultraviolet (UV) lasers for illumination.
• Laser coherence and sample surface roughness cause speckle noise, degrading image quality.

Purpose of the Study:

• To develop a novel speckle noise suppression method for coherent laser-based microscopic imaging.
• To improve image quality in laser microscopy without requiring annotated data.

Main Methods:

• A Cycle Generative Adversarial Network (CycleGAN) was trained using statistical physical modeling and image gradient discrepancy.
• Physical constraints were incorporated to capture speckle noise mechanisms.
• The method was designed for unsupervised learning, eliminating the need for clean, paired data.

Main Results:

• The proposed method effectively suppressed speckle noise in high-resolution laser microscopy.
• It significantly outperformed traditional filtering methods and unsupervised Convolutional Neural Networks (CNNs).
• The approach demonstrated superior denoising performance and training efficiency.

Conclusions:

• The integrated physical constraints successfully guided CycleGAN training for speckle noise reduction.
• The method offers a robust solution for enhancing laser microscopic images.
• The framework has potential applications in other coherent laser imaging modalities.