In order to enhance the intelligent discrimination of true versus false fracture-vug features in electrical imaging logs from the Chengdao buried hill belt, the Geo-Swin-Unet model was constructed by integrating lithology modulation, graph semantic enhancement, multi-scale attention fusion, and geological constraints. The results show that the model achieves a segmentation accuracy of 88% and a boundary match of 85% on a complex fracture-vug dataset, surpassing existing methods. High geological consistency is demonstrated, with fracture-orientation alignment to regional stress over 85% and a pore-fracture spatial association index above 80%, while a false positive rate below 5% is maintained. It is concluded that the Geo-Swin-Unet model proves effective for identifying multi-scale geological features in noisy logging environments, offering a reliable approach for evaluating deep and complex reservoirs.