摘要
为提升复杂场景下卷积网络压接管检测精度和速度,减小模型,提出一种基于小波分析和改进轻量化神经网络的野外复杂背景下压接管视觉检测方法,使用小波分析去除图像噪声,主干网络使用轻量化GhostNet模型,并引入全维动态卷积增强主干特征提取能力,使用深度可分离卷积降低模型复杂度,嵌入改进的卷积块的注意力模块(convolutional block attention module,CBAM),关注重点特征提高模型精度,使用K-means++算法聚类锚框尺寸并线性变换,加快目标框收敛速度,使用CIoU-NMS提高检测速度与精度。实际检测结果表明,与YOLOv4模型相比改进YOLOv4轻量化模型大小大幅降低了199.7 MByte,精度仅损失2.98%,且检测速度提升了3.4 Hz,达33.9 Hz,边缘部署效能指标更优,因此,改进轻量化网络模型在检测精度、模型大小和检测速度之间达到最佳平衡。最后,野外复杂背景多场景下的检测效果也表明算法能很好满足工程实际任务中的检测需求,具有较好工程实用性。
In order to improve the detection accuracy and speed of convolutional networks,and reduce the model under miscellaneous scenes,a visual detection method of press connection pipe under complex field background based on wavelet analysis and improved lightweight neural network is proposed.The image noise is removed using wavelet analysis and the lightweight GhostNet model is used in the backbone network.In addition,omni-dimensional dynamic convolution is introduced to enhance the backbone feature extraction capability,depth separable convolution is used to reduce the model complexity,improved convolutional block attention module(CBAM)is embedded to improve the model accuracy,and K-Means++algorithm is used to cluster the anchor frame size and linear transformation to accelerate the convergence of the target frame.CIoU-NMS is used to improve detection speed and accuracy.The actual test results show that compared with YOLOv4 model,the improved YOLOv4 lightweight model size is significantly reduced by 199.7 MByte with the accuracy is only lost by 2.98%,the detection speed is increased by 3.4 Hz to 33.9 Hz and the edge deployment efficiency index is better.Therefore the improved lightweight network model achieves the best balance among detection accuracy,model size and detection speed.Finally,the detection results in the field with complex background and multiple scenes also show that the algorithm can well meet the detection requirements in practical engineering tasks,and has sound engineering practicability.
作者
邹德华
张宏伟
乔磊
赵丽媛
江维
ZOU Dehua;ZHANG Hongwei;QIAO Lei;ZHAO Liyuan;JIANG Wei(State Grid Hunan Ultra-High Voltage Transmission Company,Changsha 420100,China;Hunan Province Key Laboratory of Intelligent Live Working Technology and Equipment(ROBOT),Changsha 420100,China;Live Inspection and Intelligent Operation Technology State Grid Corporation Laboratory,Changsha 420100,China;School of Mechanical Engineering and Automation,Wuhan Textile University,Wuhan 430073,China)
出处
《湖南电力》
2024年第1期77-84,共8页
Hunan Electric Power
基金
智能带电作业技术及装备湖南省重点实验室开放课题资助项目(2022KZD1003)
国网湖南省电力有限公司科技项目(5216AJ21N005)。
关键词
压接管
复杂背景
改进卷积网络
轻量化
全维动态卷积
pressure connection pipe
complex background
improved convolutional network
lightweight
omni-dimensional dynamic convolution(ODConv)
