Digital rock physics is an effective approach for predicting the petrophysical properties of shale. While imaging technologies struggle to achieve both high resolution and wide field of view (FoV) simultaneously, it is challenging to accurately characterize the extremely heterogeneous structures and nano-scale pores enriched features. Deep learning-based super-resolution has been increasingly developed to alleviate the trade-off between resolution and FoV. Given the difficulty of obtaining real-world low-resolution (LR) and high-resolution (HR) image pairs, along with the complexity of multiple minerals in shale, in this work, we proposed a multiscale reconstruction and multi-mineral segmentation approach based on cycle-consistent generative adversarial network, i.e., SRCycleGAN and SegCycleGAN. SRCycleGAN is trained with unpaired LR and HR images, enhancing LR image quality, while SegCycleGAN segments HR grayscale images into multiple mineral phases for detailed microstructure analysis. Note that the training images for SegCycleGAN are composed of paired HR grayscale images and labeled images segmented manually. Evaluations of the trained CycleGAN models, both visual and quantitative, include peak signal-to-noise ratio (PSNR), entropy, and microstructure analysis. The results demonstrate that SRCycleGAN significantly enhances the image quality of the shale porous media, thus achieving multiscale fusion and SegCycleGAN enables a higher multi-mineral segmentation accuracy. This framework aims to facilitate better characterization and modeling of multiscale and multi-mineral pore structures in shale, which can be easily adapted to various heterogeneous porous media, supporting applications in geo-energy development, CO2 utilization and sequestration, and underground hydrogen storage.