Objective: To explore the influence of 18F-FDG activity on PET/CT image quality and establish a PET image quality restoration method to achieve radiation protection for patients. Methods: This study proposed a global feature recognition low-activity PET image quality restoration deep learning network model, (i.e., SwinUNetR-GAN) based on Swin Transformer. 124 patients’ PET/CT images who underwent 18F-FDG PET/CT scanning in 2024 were randomly selected to explore the law of PET image quality degradation under different activities of 18F-FDG, and the low-activity PET images’ quality restoration are achieved based on the SwinUNetR-GAN network. Results: As the activity of 18F-FDG decreases, the SUVmean and SUVmax in normal tissues and tumor lesions in the PET images show an increasing trend. When the activity of 18F-FDG is reduced to 10% of the current clinical activity, the SUVmean of the tumor lesion increases by about 1.1 times, while the SUVmax increases by about 2 times. In addition, the use of the SwinUNetR-GAN network can reduce the noise of 10% activity PET images, MAE decreases from 0.21 to 0.15, and Rdev decreases from 0.33 to about 0.25. Conclusion: This study clarifies the change law of quantitative parameters of normal tissues and tumor tissues in patient PET images under activity of 18F-FDG, and then, SwinUNetR-GAN network dedicated to low 18F-FDG activity PET image quality restoration was proposed, which could achieve the disease diagnosis performance of PET images while reducing the radiation dose deposited in patient.