针对输电线路异物检测中存在背景干扰、图像分辨率低且异物尺度变化大等问题,提出了一种基于改进YOLOv7的输电线路异物检测模型.首先,通过空间深度卷积(space to depth conconvolution,SPD-Conv)和多维协作注意力(multidimensional collaborative attention,MCA)机制构造新的骨干网络,加强模型对低分辨率图像特征提取及抑制背景干扰的能力,同时增加对小目标异物的关注度.其次,使用幻影卷积(ghost convolution,Ghost-Conv)改进高效分层聚合网络(efficient layer aggregation network,ELAN)的输出部分,大幅降低模型的计算量.最后,基于可伸缩交并比(scalable intersection over union,SIoU)优化损失函数,进一步提高模型的训练速度和鲁棒性.实验结果表明,所提模型在输电线路异物检测数据集上平均精度均值(mean average precision,mAP)达到95.98%,高于其他主流对比模型,同时每秒帧数(frames per second,FPS)达到64,满足输电线路异物的实时性检测.
Foreign Object Detection Model of Transmission Line Based on Improved YOLOv7
Aiming at the problems of background interference,low image resolution,and large scale variations of foreign objects in the detection of foreign objects on power transmission lines,a foreign object detection model of power transmission line based on improved YOLOv7 is proposed.Firstly,a new backbone network is constructed through space to depth conconvolution(SPD-Conv)and multidimensional collaborative attention(MCA)mechanism to enhance the model's ability to extract features from low-resolution images and to suppress background interference,thus the attention to small foreign objects is increased.Secondly,the output part of the efficient layer aggregation network(ELAN)module is improved by using ghost convolution(Ghost-Conv)to significantly reduce the model's computational complexity.Finally,based on the scalable intersection over union(SIoU)optimized loss function,the model's training speed and robustness are further improved.Experimental results show that the proposed model achieves a mean aver-age precision(mAP)of 95.98%on the power transmission line foreign object detection dataset,which is higher than other mainstream comparative models.At the same time,the frames per second(FPS)reaches 64 to meet the real-time detection require-ments of foreign objects of power transmission lines.
foreign object of transmission lineYOLOv7MCAsmall objectSPDGhost-Conv
国家科技期刊平台
NSTL
万方数据
