| High-resolution(HR)astronomical images are essential for scientific research,cosmic exploration,astronomy,and the development of physics.Astronomers can use high-resolution astronomical images to study the interactions between stars and the interstellar medium in galaxies and to explore the formation and evolution of stars and planets.However,due to the constraints of astronomical observation,astronomical images are susceptible to atmospheric refraction,light pollution,and electronic noise,which degrade the image quality of astronomical images and affect the identification and analysis of the observed targets.Therefore,it is especially important to reconstruct low-resolution(LR)astronomical images by the super-resolution(SR)reconstruction technique to improve the resolution of astronomical images.To this end,this paper proposes a new super-resolution reconstruction algorithm for astronomical images,and the main research includes the following two aspects.(1)To address the problem of low-resolution astronomical image imaging,this paper proposes a Self-Attention Super-Resolution Generative Adversarial Network(SASRGAN)based SRGAN astronomical image super-resolution reconstruction algorithm.The algorithm uses SRGAN as the benchmark model,and first introduces a self-attention in the network to capture more global dependencies and increase the network depth to better represent highfrequency features.Secondly,the BN layer in the generative network is removed in order to improve the stability and efficiency of training.Finally,to better handle outliers and improve SR performance,Charbonnier loss is introduced to replace the MSE loss in the loss function,and full-variance loss is added to maintain the smoothness of the image and suppress the generation of image artifacts.The experimental results show that SASRGAN outperforms the benchmark model SRGAN in all evaluation indexes,and also shows excellent SR performance in comparison with the reconstruction of other classical image super-resolution algorithms,and the reconstructed astronomical images have better perceptual effects.(2)To address the problems of high computational complexity and low training efficiency of SASRGAN,this paper proposes an enhanced Self-Attention Super-Resolution Generative Adversarial Network(ESASRGAN)based on interlaced sparse self-attention and a mixture of augmentation is proposed for astronomical image super-resolution reconstruction.Firstly,the interleaved sparse self-attention is introduced to optimize the self-attention module in the network to reduce the complexity and GPU occupation of the model without affecting the SR performance and improve the training efficiency of the model.Second,a mixture of augmentation technique is used to increase the diversity of data samples,improve the stability and generalization ability of the network,and reduce the risk of model overfitting.The experimental results show that ESASRGAN significantly improves the performance efficiency of the model,the GPU occupation is reduced to 42.5% of the original,the FLOPs are reduced to 52.9% of the original,the inference speed is nearly doubled,and it outperforms SASRGAN in all evaluation indexes,and the reconstructed astronomical images are more realistic and clear.In summary,the super-resolution reconstruction algorithm of astronomical images proposed in this paper can effectively reconstruct low-resolution astronomical images and generate high-resolution astronomical images with rich and clear details. |
