DSR-Net: Dynamic Star Map Denoising Algorithm Based on Deep Reinforcement Learning

As astronomical observation technology continues to progress, obtaining high-quality star maps provides us with valuable opportunities to explore the universe. However, The acquired star maps are often affected by various random noises, including speckle noise, Poisson noise, Impulse noise, Thermal noise,Reynolds noise, and Gaussian noise,etc. These noises degrade the image quality and limit the efficiency of scientific research. Traditional denoising methods are often limited in their effectiveness when faced with such complex noise and lack the ability to model the temporal features of dynamic star maps, making it difficult to handle the sparsity and complex background of dynamic star maps. Therefore, this paper introduces DSR-Net, a deep reinforcement learning-based dynamic star map denoising algorithm. The algorithm combines Convolutional Gated Recurrent Units (ConvGRU) and Region-based Reward Convolution (Rrc) modules, enabling it to capture the dynamic changes of star maps, effectively remove noise, and preserve important details in the star maps. Experimental results show that DSR-Net outperforms traditional denoising methods on multiple real dynamic star map datasets, providing an effective solution for improving the quality of star maps.