【摘要】 针对绝缘子缺陷目标区域较小、部分缺陷特征相似，从而导致检测精度较低的问题，提出了一种特征表示增强模型（FLDM-YOLO）。该模型基于FasterNet重构特征提取网络并且结合大核可分离注意力（LSKA）设计了SPPF-LSKA模块，增强了对目标的特征提取能力；以重参数化技术为基础，提出了C2f-DBB模块，处理目标缺陷特征相似的问题；在边界框回归阶段使用MPDIoU作为损失函数，使得模型更加关注高质量锚框。实验结果表明，FLDM-YOLO模型在保证一定检测速度的前提下，mAP为91.3%，较YOLOv8模型提高了4.2%，可有效应用于实际的巡检工作。更多还原

【Abstract】 To solve the problems posed by the small object area of insulator defects and the similarity in some defect features, which lead to low detection accuracy, a feature representation enhancement model(FLDM-YOLO) is proposed. This model is based on the reconstruction of the feature extraction network using FaterNet and combined with large kernel separable attention(LSKA) to design the SPPF-LSKA module, which enhanced the feature extraction ability of the object. Based on reparameterization technology, the C2f-DBB module is proposed to address the issue of similar defect features in objects. Utilizing MPDIoU as the loss function in the bounding box regression stage shifted the model’s focus toward high-quality anchor boxes. Experimental results indicated that the FLDM-YOLO model achieved a mean average precision(mAP) of 91. 3%, 4. 2% higher than the YOLOv8 model while maintaining a reasonable detection speed, making it suitable for actual inspection work.更多还原