Abstract: To comprehensively understand the regularities of solar activity, it is necessary to continuously collect high-quality data covering multiple solar activities cycles over a long period of time. Solar magnetograms are the most important data for studying solar activity. Continuous, long-term, and high spatial resolution solar magnetograms can provide more detailed information on solar magnetic field evolution, which is helpful for more accurate prediction of solar activity and space weather events. Therefore, this paper proposes a deep learning-based super-resolution algorithm to perform super-resolution on MDI (Michelson Doppler Imager) magnetograms to achieve the same resolution as HMI (Helioseismic and Magnetic Imager) magnetograms, thus obtaining a high-quality solar magnetogram database covering nearly two solar activity cycles. To improve the performance of the model, this paper introduces an attention mechanism in the deep neural network, which effectively guides the network to enhance effective feature information by learning attention weight maps in the channel and spatial dimensions. In addition, this paper uses uncertainty loss as the loss function for model training, which can assign greater weight to textures and edges with magnetic field changes, without increasing the network parameters and computational complexity. The experimental results show that the proposed algorithm significantly improves the quality of super-resolution solar magnetograms and achieves the best results in peak signal-to-noise ratio (PSNR: 33.3168), Structure Similarity Index Measure (SSIM: 0.8754), correlation coefficient (CC: 0.9323), root mean square error (RMSE: 21.8544) and other indicators.