基于交错部分卷积的高压输电线路检测方法

High-voltage Transmission Line Detection Method Based on Interleaved Partial Convolution

在输电线路无人机巡检任务中,针对基于深度学习的航拍图像中待检测目标检测精度不高和模型过大而难以部署至无人机等移动端设备的问题,提出了以YOLOv7-tiny为基础网络进行改进以实现提高检测精度并将模型轻量化的方法。首先,该文设计了一种交错部分卷积(interlace partial convolution,IPConv),并利用其构建IP1-ELAN、IP2-ELAN模块作为网络的特征提取模块,使其能有效减轻模型中通道冗余问题,并大幅度减少模型的参数量和浮点数;其次,在骨干网络最后一层中融合高效多尺度注意力机制(efficient multi-scale attention,EMSA)以实现跨通道交互,增强目标区域特征提取能力;最后,融合快速空间金字塔池化及跨阶段空间通道(spatial pyramid pooling faster,cross stage partial channel,SPPFCSPC)模块,进一步增强特征提取能力,提升模型检测性能。通过实验验证,该文方法在输电线路巡检数据集中模型参数量和浮点数分别仅为3.79M,8.4G,检测精度为85.8%。综合性能优于目前常用的检测算法,能够基本满足部署至无人机端进行检测任务。

In the transmission line UAV inspection task,to address the problems of low detection accuracy of targets to be detected in deep learning-based aerial images and the model being too large and complicated to be deployed to mobile devices such as UAVs,a method of improving YOLOv7-tiny as the base network to achieve improved detection accuracy and light_weight of the model is proposed.Firstly,this paper designs an Interlace Partial Convolution(IPConv)and uses it to construct IP1-ELAN and IP2-ELAN modules as the feature extraction module of the network,so that it can effectively alleviate the problem of channel redundancy in the model and substantially reduce the number of parameters and floating points in the model;secondly,in the last layer of backbone Secondly,Efficient Multi-Scale Attention(EMSA)is integrated in the previous layer of the network to achieve cross-channel interaction and enhance the feature extraction capability of the target region;finally,SPPFCSPC(Spatial Pyramid Pooling Faster,Cross Stage Partial Channel)module to further enhance the feature extraction capability and improve the model detection performance.Through experimental verification,the number of model parameters and floating points in this method’s transmission line inspection dataset are only 3.79M and 8.4G,respectively,and the detection accuracy is 85.8%.The comprehensive performance is better than that of the current commonly used detection algorithms,and can basically meet the deployment to the UAV end for detection tasks.