摘要
针对形态学和色彩学图像处理算法检测电线缺陷类型少、效率低的问题,为提高电线缺陷的检测正确率,提出了一种高效率、高准确度的高压电线缺陷图像识检测方法。由于电线缺陷类型根据导线截面积变化情况来划分,故选取更快速的基于区域卷积神经网络(Faster R-CNN)对电线缺陷进行检测与分类。根据检修导则将电线缺陷分为5种单类和2种混合类,研究了不同的网络模型对电线缺陷检测的正确率和识别帧率,并在实验中对数据集进行变换,通过旋转图像和加入正态分布的高斯雪花进一步提升检测效果。通过采集到的电线缺陷图像的测试,Faster R-CNN能够检测分类高压电线缺陷,加入数据变换对于电线缺陷检测的有效性和可靠性都有了提升,识别帧率为119 ms,均值平均精度(mAP)为94%,提高了5个百分点。
For the problem of less defect types and low efficiency of wire detection by morphological and color image processing algorithms,in order to improve the detection accuracy of wire defects,a high efficient and high accurate method for detecting defects of high voltage wires was proposed.Since the types of wire defects were divided according to the change of the cross-sectional area of the wire,Faster Region-Based Convolutional Neural Network(Faster R-CNN)was selected to detect and classify the wire defects.According to the inspection guide,the wire defects were divided into five single types and two hybrid types.The accuracy and recognition frame rate of the wire defect detection by different network models were studied.By rotating images and adding a normal distribution of Gaussian snowflakes,the data sets were transformed in the experiment,which further enhanced the detection effect.Through the test of the collected wire defect images,Faster R-CNN can detect the types of high-voltage wire defects,and the data transformation improves the effectiveness and reliability of wire defect detection.The recognition frame rate is 119 ms.The mean Average Precision(mAP)is 94%,which is 5 percentage points higher than that without data transformation.
作者
金昊
康宇哲
齐希阳
洪榛
JIN Hao;KANG Yuzhe;QI Xiyang;HONG Zhen(Faculty of Mechanical Engineering and Automation,Zhejiang Sci-Tech University,Hangzhou Zhejiang 310018,China;School of Information Science and Technology,Zhejiang Sci-Tech University,Hangzhou Zhejiang 310018,China)
出处
《计算机应用》
CSCD
北大核心
2019年第S02期97-102,共6页
journal of Computer Applications
基金
国家级大学生创新创业训练计划项目(201810338001)
关键词
高压电线缺陷
区域卷积神经网络
数据变换
检测正确率
识别帧率
high-voltage wire detection
Region-based Convolutional Neutral Network(R-CNN)
data transmission
defection efficiency
defection frame rate
