Tight glutenite, as key carriers of unconventional oil and gas resources, have become a research focus in the energy field for their geological features and engineering value. Cores, direct samples of underground reservoirs, need accurate splitting and characterization to understand fluid flow and predict production. But this process is challenged by complex pore structures, strong heterogeneity, and low porosity and permeability. Advanced architectures like U-Net and U-Net++, widely used in the industry, have problems like insufficient global context modeling, reliance on large-scale labeled data, and high computational complexity. So, they can't efficiently and robustly segment tight glutenite with extremely imbalanced matrix and pore distributions. This study offers a solution based on the SegFormer framework. Its hierarchical Transformer encoder captures multi - scale global features. A lightweight MLP decoder combines features across resolutions. Focal Loss dynamically adjusts sample weights to ease the imbalance between pore and matrix distributions. Experiments show the new model reaches an IoU of 91.97% and 97.71% on tight sandstone and conglomerate datasets, a 3.03% improvement over U-Net. Further ablation experiments prove Focal Loss optimizes the conglomerate pore segmentation IoU from the failure state under cross - entropy loss to 97.71%, significantly reducing segmentation model bias from lithological heterogeneity. This is the first application of SegFormer in core analysis, offering an efficient way for precise pore - network modeling in tight reservoirs.