针对无人机航拍绝缘子图像具有小目标、多尺度、多形态、朝向任意以及缺陷特征不显著等特点，提出一种改进ReDet的定向检测算法。首先，将旋转等变注意力模块(Rotation-equivariant CBAM, ReCBAM)集成到主干网络中，增强目标区域的聚焦能力；其次，结合空洞卷积组、自门控激活函数以及自适应特征融合模块，构建了一种新型旋转等变递归特征金字塔，提供更强大的多尺度、多形态特征建模能力；最后，将RoI检测器扩展为四阶段混合级联结构，同时引入上下文信息融合策略与改进的损失函数，进一步精细化目标区域的特征表达。其中，改进的损失函数结合KL散度、Focal Loss机制以及动态IoU权重，建立旋转检测框的几何-空间动态联合优化机制，提升了定位精度。在自建数据集上的实验结果表明，在旋转框标注下，本文算法mAP值达到95.06%，较基础ReDet提升了3.01个百分点，在复杂背景下的旋转小目标检测中表现优异。且模型参数量仅为34.57MB，适合部署于无人机平台；同时在水平框标注下mAP达到95.56%，较基础ReDet提升了1.57个百分点，验证了算法的良好泛化能力。

To address the challenges of insulators in UAV aerial images, which exhibit characteristics such as multi-scale, diverse morphologies, arbitrary orientations, and inconspicuous defect features, an improved ReDet-based orinented detection algorithm is proposed. First, a Rotation Equivariant Attention Module (ReCBAM) is integrated into the backbone network to enhance the focus on target regions. Second, a novel rotation-equivariant recursive feature pyramid is constructed using dilated convolution group, self-gated activation functions, and an adaptive feature fusion module, providing stronger multi-scale and multi-morphological feature modeling capabili-ties. Finally, the RoI detector is extended to a four-stage hybrid cascade structure, incorporating contextual information fusion and an improved loss function to refine the feature representation of target regions. The improved loss function combines KL divergence, the Focal Loss mechanism, and an IoU-based dynamic weighting mechanism, establishing a geometric-spatial dynamic joint optimi-zation mechanism for rotated bounding boxes, effectively improving the localization accuracy. Experimental results on a self-built dataset show that the proposed algorithm achieves a mAP of 95.04% under oriented bounding box annotations, representing a 3.01% improvement over the baseline ReDet, and demonstrates superior performance in detecting small, rotated targets under complex backgrounds. Additionally, the model's parameter size is only 34.57 MB, making it suitable for deployment on UAV platforms. Fur-thermore, under horizontal bounding box annotations, the algorithm achieves a mAP of 95.56%, with a 1.57% improvement over the baseline ReDet, validating the algorithm's strong generalization capability.