Multi-Frame Star Image Denoising Algorithm Based on Deep Reinforcement Learning and Mixed Poisson-Gaussian...

Area of Science:

• Astronomy
• Image Processing
• Machine Learning

Background:

• Star images often contain mixed Poisson-Gaussian noise, which is challenging to remove using standard Maximum Likelihood Estimation (MLE).
• The complexity of the likelihood function in MLE hinders effective noise suppression in astronomical imaging.

Purpose of the Study:

• To develop an accurate star image restoration method for mixed Poisson-Gaussian noise.
• To integrate Maximum Likelihood Estimation (MLE) with advanced machine learning for improved denoising.

Main Methods:

• A reinforcement learning (RL) algorithm was employed, utilizing the mixed Poisson-Gaussian likelihood function as its reward function.
• The agent navigated a Markov Decision Process (MDP) to generate restored images maximizing the likelihood function.
• Hyperparameter tuning was performed through simulated experiments to optimize denoising performance.

Main Results:

• The RL-based algorithm demonstrated superior performance in suppressing mixed Poisson-Gaussian noise compared to traditional MLE.
• The proposed method outperformed a Deep Convolutional Neural Network (DCNN) based denoising technique.
• Experimental results validated the effectiveness of the RL approach for star image restoration.

Conclusions:

• Reinforcement learning offers a powerful framework for tackling complex noise models in star image processing.
• The developed RL method provides a more accurate and effective solution for denoising compared to existing techniques.
• This approach advances the field of astronomical image restoration by leveraging state-of-the-art machine learning.