李利荣, 梅冰, 戴俊伟, 毛锐, 童庆刚, 时愈. 基于多尺度轻量化卷积的高压输电线路巡检[J]. 高电压技术, 2024, 50(12): 5269-5280. DOI: 10.13336/j.1003-6520.hve.20231754
引用本文: 李利荣, 梅冰, 戴俊伟, 毛锐, 童庆刚, 时愈. 基于多尺度轻量化卷积的高压输电线路巡检[J]. 高电压技术, 2024, 50(12): 5269-5280. DOI: 10.13336/j.1003-6520.hve.20231754
LI Lirong, MEI Bing, DAI Junwei, MAO Rui, TONG Qinggang, SHI Yu. High-voltage Transmission Line Inspection Based on Multi-scale Lightweight Convolution[J]. High Voltage Engineering, 2024, 50(12): 5269-5280. DOI: 10.13336/j.1003-6520.hve.20231754
Citation: LI Lirong, MEI Bing, DAI Junwei, MAO Rui, TONG Qinggang, SHI Yu. High-voltage Transmission Line Inspection Based on Multi-scale Lightweight Convolution[J]. High Voltage Engineering, 2024, 50(12): 5269-5280. DOI: 10.13336/j.1003-6520.hve.20231754

基于多尺度轻量化卷积的高压输电线路巡检

High-voltage Transmission Line Inspection Based on Multi-scale Lightweight Convolution

  • 摘要: 无人机搭载轻量化模型进行高压输电线路巡检对于保障输电系统的安全运行至关重要,针对目前轻量化模型中特征提取不够丰富导致高压输电线路巡检效果较差的问题,该文提出了一种多尺度轻量化卷积(multi-scale lightweight convolution,MSLConv),其进行了不同尺度与深度的特征融合,并实现了网络轻量化。同时在混洗注意力(shuffle attention,SA)中提出一种新的更具适应性的通道混洗方法,并将MSLConv与SA相结合设计了MSLConvSA1与MSLConvSA2特征提取模块,提高了MSLConv的通道利用率。在处理样本输出时,提出了一种基于样本的双重归一化与激活的组合方法,该方法能在通道数较少的轻量化模型中对通道特征进行更细粒度的调整,且能更好地建立样本间的特征关系。最后以YOLOv5n为基线进行一系列改进,得到了MSV1与MSV2两种模型。经实验证明,相对于基线,这两种模型的参数量分别减少了43.8%和28.6%,计算量分别减少了38.1%和23.8%。同时,mAP@.5(IoU阈值为0.5时的平均检测精度值)分别提升了5.2%和5.4%,mAP@.5:.95(IoU阈值为0.5~0.95,取步长为0.05时的平均检测精度值)分别提升了3.3%和4.7%。

     

    Abstract: The use of unmanned aerial vehicles equipped with lightweight models for high-voltage power transmission line inspections is crucial for ensuring the safe operation of power transmission systems. Currently, insufficient feature extraction in lightweight models may lead to inferior effectiveness in high-voltage power transmission line inspections, thus this paper proposes a multi-scale lightweight convolution (MSLConv). MSLConv fuses features at different scales and depths while achieving network lightweight. In the shuffle attention (SA) module, a novel and more adaptive channel shuffling method is introduced. The MSLConv is combined with SA to design two feature extraction modules, MSLConvSA1 and MSLConvSA2, thereby enhancing the channel utilization of MSLConv. When processing sample output, this paper proposes a combined method based on sample-based dual normalization and activation. This method allows for finer adjustments of channel features in lightweight models with fewer channels and a better establishment of feature relationships between samples. Finally, this paper conducts a series of improvements based on the YOLOv5n baseline, resulting in two models: MSV1 and MSV2. The experimental results demonstrate that, compared to the baseline, MSV1 and MSV2 can reduce parameter quantities by 43.8% and 28.6%, and reduce computational loads by 38.1% and 23.8%, respectively. Additionally, mAP@.5(average detection accuracy value at an IoU threshold of 0.5) is improved by 5.2% and 5.4%, and mAP@.5:.95(average detection accuracy value when the IoU threshold is 0.5~0.95 and the step size is 0.05) is improved by 3.3% and 4.7% for MSV1 and MSV2, respectively.

     

/

返回文章
返回