Low-resolution and blurry computed tomography (CT) images of porous media will lose useful details and may result in inaccurate evaluation of physical properties. Therefore, resolution enhancement and deblurring of CT images of porous media are of great significance to understand the structural and physical properties. In this paper, a novel Super Resolution Generative Adversarial Network based on High-Resolution Representation Learning (HRGAN) is developed to improve μ-CT image quality, which has been tested on Estaillades carbonate, naturally fractured coal, Savonnières carbonate, Massangis Jaune carbonate and Bentheimer sandstone. The PSNR (peak signal-to-noise ratio), SSIM (structural similarity index) and AP (area porosity) as well as FD (fractal dimension) are employed to compare and examine the performance of different network structures. It has been found that the reconstructed textures by the present HRGAN are more precise than that with traditional Super Resolution Generative Adversarial Network based on ResNet Generator (SRGAN). Compared with the best performance of SRGAN, the training loss and test loss of HRGAN with multi-resolution subnetworks in parallel are lowered by 17.24% and 4.09%, and the PSNR, SSIM, AP and FD are enhanced by 6.42%, 11.25%, 22.6% and 39.5%, respectively. Furthermore, the generalization of HRGAN is evaluated by measured μ-CT images of Ottawa sand. The proposed HRGAN indicates evident advantages in precision and generalization of resolution enhancement and deblurring of μ-CT images.