摘要
针对现场中采集的绝缘子图像存在目标图像大小尺度不一,以及拍摄角度所造成的目标图像相互遮挡等因素而导致误检或漏检等问题,提出了一种改进的基于卷积神经网络的绝缘子图像检测方法。采用轻量化的ZF网络实现特征提取;确定优化的锚窗比例提升目标图像的检测精度;对NMS后处理算法进行了改进,提出多阶段的惩罚因子算法,适应于多尺度、多比例、绝缘子重叠遮挡等复杂情况。实验结果表明,改进后的Faster R-CNN的检测方法将AP由0.797 7提高到了0.905 8,显著地提升了绝缘子目标图像的检测精度,降低了绝缘子的漏检和误检的概率。
Aiming at the problem of the misdetection or omission caused by the different sizes and scales of target images and the mutual occlusion of target images caused by the shooting angle,an improved detection method of insulator images based on convolutional neural network is proposed in this paper.Firstly,the lightweight ZF network is adopted to achieve feature extraction,and then,the optimized anchor window ratio is determined to improve the detection accuracy of the target image.Finally,the NMS post-processing algorithm is improved,and a multi-stage penalty factor algorithm is proposed,which is suitable for complex situations such as multi-scale,multi-ratio and overlapping insulators.Experimental results show that the improved detection method of Faster R-CNN increases the AP from 0.797 7 to 0.905 8,which significantly improves the detection accuracy of insulator target image,and reduces the probability of insulator omission and misdetection.
作者
吴君鹏
唐少博
李相磊
张师
Wu Junpeng;Tang Shaobo;Li Xianglei;Zhang Shi(Northeast Electric Power University,Jilin 132000,Jilin,China)
出处
《电测与仪表》
北大核心
2022年第5期116-122,共7页
Electrical Measurement & Instrumentation
基金
吉林省教育厅资助项目(JJKH20180440KJ)
吉林省科技发展计划项目(20200403075SF)。
