Astronomical Image Denoising Using AttentionGAN

Abstract

This research presents a novel approach to astronomical image denoising using AttentionGAN, a generative adversarial network architecture enhanced with attention mechanisms. Traditional denoising methods often fail to preserve fine details in astronomical images, such as stars and nebulae. Our proposed method leverages CycleGAN's unpaired learning capability with attention modules to effectively remove noise while preserving critical astronomical features. Experimental results demonstrate significant improvements in both quantitative metrics and visual quality compared to existing approaches.

1. Introduction

Astronomical imaging presents unique challenges for image processing. When capturing faint celestial objects, astronomers must deal with various sources of noise:

• Photon shot noise - Inherent quantum fluctuations in light detection
• Read noise - Electronic noise from sensor readout
• Thermal noise - Heat-induced signals in detectors
• Cosmic rays - High-energy particle impacts

Traditional denoising methods like Gaussian filtering or median filtering often blur fine details - precisely the features astronomers care about most.

2. Methodology

2.1 CycleGAN Architecture

We built upon CycleGAN, which learns image-to-image translations without requiring paired training data. This is crucial for astronomical imaging where obtaining matched noisy/clean image pairs is extremely difficult.

2.2 Attention Mechanisms

The key innovation is integrating attention layers into the generator:

1. Self-Attention Module - Enables the model to focus on relevant spatial features across the entire image, helping distinguish between noise and actual astronomical features

2. Channel Attention - Emphasizes important feature channels, allowing the network to prioritize certain types of astronomical structures

3. Residual Attention - Helps preserve fine details while learning complex noise patterns

2.3 Loss Functions

The model uses multiple loss functions:

• Adversarial Loss - Ensures realistic output
• Cycle Consistency Loss - Maintains astronomical integrity
• Perceptual Loss - Preserves fine details using pre-trained VGG features

3. Experimental Results

Dataset

• Generated synthetic noisy astronomical images
• Added varying levels of Gaussian noise (σ = 10, 20, 30)
• Used high-resolution star field and nebula images

Quantitative Results

Method	PSNR	SSIM
BM3D	28.45	0.85
DnCNN	30.12	0.89
CycleGAN	31.23	0.91
AttentionGAN	33.87	0.94

Key Findings

1. Noise Reduction - Achieved 15% better noise reduction than CycleGAN
2. Detail Preservation - Maintained stellar objects with 94% SSIM
3. Adaptability - Successfully handled various noise levels

4. Visual Results

The AttentionGAN successfully:

• Removes background noise while preserving star points
• Maintains nebula structures without over-smoothing
• Handles varying noise levels without parameter tuning

5. Conclusion

AttentionGAN demonstrates that combining generative adversarial networks with attention mechanisms offers a powerful solution for astronomical image denoising. The approach successfully preserves critical astronomical details while effectively removing various noise sources.

This research was conducted as part of my Master's thesis at the University of Indonesia, combining my background in Astronomy with machine learning expertise.

---

Read the full paper on ResearchGate