针对航拍角度导致的水平矩形框中的绝缘子检测背景区域过大的问题，采用旋转目标框替代水平矩形框对绝缘子进行检测，构建Ghost-YOLOv5s网络，实现旋转绝缘子缺陷高精度、快速检测。将轻量化的GhostNet作为YOLOv5s主干网络，减少网络参数，提高检测速度；颈部网络采用双向特征金字塔结构，并引入可学习的通道权重进行自适应加权，加强网络浅层特征的有效复用，提高检测精度；头部网络预测模块引入旋转量，使预测框定位更精准，减少漏检，提高检测准确性；采用相对熵损失函数计算边界框回归损失，解决引入旋转量导致的损失值突变问题；添加第4检测层并聚类分析绝缘子数据集确定预测锚框尺寸，提高网络小目标检测精度；利用剪枝算法压缩模型体积，加快检测速度。实验结果表明：剪枝后的Ghost-YOLOv5s网络检测精度和速度分别为95.9%和78.6 帧/s，满足绝缘子缺陷检测要求。

In order to solve the problem that the background area of insulator detection in the horizontal rectangular frame becomes excessively large due to aerial shooting angles， a rotating bounding box is used instead of the horizontal rectangular frame to detect the insulator， and the Ghost-YOLOv5s network is constructed to achieve high-precision and fast detection of insulator defects using rotating bounding boxes. The lightweight GhostNet is employed as the backbone network of YOLOv5s to reduce network parameters and enhance detection speed. In the neck network， a bidirectional feature pyramid structure is adopted along with the introduction of learnable channel weights for adaptive weighting， strengthening the effective reuse of shallow features in the network and improving detection accuracy. The head network prediction module introduced rotation variables to improve the accuracy of bounding box localization， reduce missed detections， and improve detection accuracy. A relative entropy loss function is used to calculate the bounding box regression loss， addressing the issue of loss value mutation caused by the introduction of rotation variables. A fourth detection layer is added， and clustering analysis on the insulator dataset determine the sizes of the predicted anchor boxes， improving the network’s precision in detecting small targets. A pruning algorithm is applied to compress the model size， further increasing detection speed. Experimental results show that the pruned Ghost-YOLOv5s network achieve a detection precision and speed of 95.9% and 78.6 fps， respectively， meeting the requirements for insulator defect detection.