Non-Uniformity Correction of Spatial Object Images Using Multi-Scale Residual Cycle Network (CycleMRSNet)

Area of Science:

• Astronomy
• Computer Vision
• Image Processing

Background:

• Ground-based telescopes face stray light and vignetting, causing non-uniform backgrounds.
• Non-uniform backgrounds degrade signal-to-noise ratio for target tracking and recognition accuracy.
• Existing methods struggle to effectively correct these image artifacts.

Purpose of the Study:

• To propose CycleMRSNet, a novel network architecture for correcting non-uniform backgrounds in astronomical images.
• To enhance image processing capabilities using a multi-scale attention mechanism within a CycleGAN framework.
• To improve the accuracy and robustness of space object tracking and recognition systems.

Main Methods:

• Developed CycleMRSNet based on the CycleGAN framework.
• Integrated a multi-scale feature extraction module (MSFEM) in the generator.
• Embedded efficient multi-scale attention residual blocks (EMA-residual blocks) in the Resnet backbone.
• Trained the model on a small-scale dataset and tested on simulated and real images.

Main Results:

• CycleMRSNet achieved high performance metrics: PSNR 32.7923, SSIM 0.9814, and FID 1.9212 on the test set.
• The model significantly outperformed existing methods in background correction.
• Demonstrated improved focus on multi-scale information in high-dimensional feature maps.

Conclusions:

• CycleMRSNet effectively corrects non-uniform backgrounds in images from ground-based telescopes.
• The proposed architecture enhances feature extraction efficiency and attention to critical image areas.
• The method improves the overall robustness and accuracy of astronomical imaging systems.