The detection of transformer oil leakage is crucial for ensuring the safety of electric power supply. However, oil leakage from transformers and shadows formed by illuminated transformer bodies are considerably similar, with irregular shapes and variable size. Hence, it is difficult to accurately identify the oil leakage area from transformer oil leakage images. In this study, a transformer oil leakage detection model based on spatially enhanced transformer (SE-FormerSeg) was proposed. First, to enhance the ability to distinguish the features of oil leakage and shadow, the proposed spatially enhanced transformer (SE-Former) was added to the backbone network to enhance the network feature learning ability. Second, to solve the problem of irregular shape of the detection area, a multi-scale progressive atrous convolution pyramid (MSPAP) was proposed to extract the features of the detection area under various receptive fields. Finally, to further adapt to feature learning for oil leakage and shadow in different regions, a multi-stage feature fusion (MSFF) method combining high- and low-level features was proposed. Experimental results showed that the proposed SE-FormerSeg model can not only effectively distinguish between oil leakage and shadow, but also improve the detection of transformer oil leakage. The average intersection over union (66.33%), average precision (82.62%), and average recall rate (75.24%) were significantly improved compared with those of the mainstream model. In addition, comparative experiments were performed on open marine oil leakage and shadow datasets. The results demonstrated the good generalization ability of the proposed method.