Astronomical Image Quality Assessment Based on Deep Learning for Resource-constrained Environments

Abstract This paper presents a highly lightweight model for astronomical image quality assessment, named AQSA-Net, designed to address the challenges of evaluating image quality in scenarios with limited computational resources and rapid decision-making needs. With only 0.15B in computational cost and 0.67M parameters, AQSA-Net significantly reduces memory usage while enhancing inference speed and real-time processing. We construct a data set with eight quality categories based on actual astronomical images. We develop an efficient feature extraction and data processing method that integrates local and global image information, substantially reducing input resolution and training time. We optimize the AQSA-Net architecture and introduce a spatial attention unit, enabling the model to focus on key image areas, enhancing feature extraction while reducing computational overhead. AQSA-Net is compared with several classic deep convolutional neural networks, and experimental results show that AQSA-Net achieves state-of-the-art performance with minimal computational complexity and parameter count. Specifically, AQSA-Net achieves an accuracy of 97.63%, recall of 98.03%, precision of 97.79%, and F1-score of 97.91% on the test set. Additionally, to more accurately assess the quality of usable images, we construct a quantitative image quality factor and a quality grading system, providing quantifiable evaluation criteria for subsequent scientific research. Therefore, our method effectively distinguishes high-quality images from low-quality ones that may impact scientific projects. This provides a reliable automated quality assessment tool for large-scale, complex data sets requiring deep learning inspection. Furthermore, our image quality evaluation could support the assessment of scientific observation data.