针对目前变压器外观缺陷检测精度低无法兼顾效率的问题，基于YOLOv11n进行改进。首先设计一种倒置残差注意力机制(iEMA),有效利用长距离依赖关系，并动态地计算特征权重，旨在提升变压器外观缺陷检测的准确性和鲁棒性；其次利用注意力(HCA)模块改进YOLOv11中的C2PSA,在保留局部空间信息的精准性的同时，提升全局特征的丰富性，并与iEMA形成多尺度特征提取-跨通道特征强化的路径；最后针对目标检测任务里尺寸变化较大的物体，利用Soft＿NMS对Inner＿MPDIoU改进，为细小的缺陷检测提供更准确的度量方法。实验结果表明，改进算法相比YOLOv11n, mAP@0.5提升了5.1%,召回率提升2.4%,且算法检测速度达到66.6 FPS,满足实时检测的需求。相比其他模型，改进后的YOLOv11对于变压器外观缺陷检测具有一定的优异性。

In response to the current problem of low accuracy and inability to balance efficiency in detecting appearance defects in transformers, this article proposes improvements based on YOLOv11n. Firstly, design an inverted residual attention mechanism(iEMA) that effectively utilizes long-range dependencies and dynamically calculates feature weights, aiming to improve the accuracy and robustness of transformer appearance defect detection; Secondly, utilizing the parallel attention module to improve C2PSA in YOLOv11, while preserving the accuracy of local spatial information, enhances the richness of global features; Finally, for objects with significant size changes in object detection tasks, the Soft-NMS algorithm is combined with Inner＿MPDIOU to address the limitations of traditional loss functions in detecting irregular objects and objects with significant size changes, providing a more accurate measurement method. The experimental results show that compared to YOLOv11n, the improved algorithm proposed in this paper, mAP@0.5 improved by 5.1%, recall rate increased by 2.4%, and algorithm detection speed reached 66.6 FPS, meeting the requirements of real-time detection. Compared with other models, the improved YOLOv11 has certain advantages in detecting appearance defects of transformers.