基于行列特征关联的通用倾斜绝缘子检测算法

A General Inclined Insulator Detection Algorithm Based on Information Sharing

摘要:

针对现有绝缘子检测算法难以适配多种边缘终端检测设备、倾斜绝缘子目标检测框贴合度不足的问题，开展通用倾斜绝缘子检测算法研究。采用多规格Efficient-Net网络替换YOLOv8s算法的主干特征提取网络，提出行列级共享关联的非局部结构以强化特征表征、解决特征稀疏性与有效性问题，引入目标框四个顶点的偏移量预测实现倾斜绝缘子精准定位；结合实际场景需求，构建更贴合应用的APR评价指标。在中国电力线绝缘子数据集（CPLID）上的实验表明，该算法检测倾斜目标时APR指标较基准算法提升1.6%，检测速度缩短至基准算法的56.53%，计算复杂度降至基准算法的46.42%，可适配多种硬件设备的实际应用需求。

Abstract:

Aiming at the problems that existing insulator detection algorithms have difficulties adapting to various edge terminal detection devices and exhibit insufficient fitting of detection boxes for inclined insulator targets, this study conducts research on a general inclined insulator detection algorithm. A multi-specification EfficientNet network is adopted to replace the backbone feature extraction network of the YOLOv8s algorithm; a row-column shared associated non-local structure is proposed to enhance feature representation and address the issues of feature sparsity and effectiveness; meanwhile, offset prediction for the four vertices of the target box is introduced to achieve accurate positioning of inclined insulators. Combined with the requirements of practical scenarios, a more application-oriented evaluation metric Average Precision-Recall Composite Metric (APR) is developed. Experiments on the Chinese Power Line Insulator Dataset (CPLID) demonstrate that when detecting inclined targets, the APR metric of the proposed algorithm is 1.6% higher than that of the benchmark algorithm, the detection speed is reduced to 56.53% of the benchmark algorithm, and the computational complexity is reduced to 46.42% of the benchmark algorithm. This algorithm can meet the practical application requirements of various hardware devices.