摘要
针对Faster RCNN算法在电力设备状态图像处理过程中生成的Anchor与目标设备不匹配而导致的设备识别率降低的问题,提出了一种基于连通域预处理的Faster RCNN的改进模型。依据像素值相近及位置相邻原则构建连通域,将连通域的长宽比作为输入信息对RPN(region proposal network,RPN)的Anchor面积框进行修正,以此提高Anchor boxes与目标设备的匹配度。实验结果表明:识别区域与目标设备面积的进一步匹配避免了多余锚框网络计算量,该方法对目标设备误识别率由改进前的2.2%下降到1.9%,降幅达14%,同时其计算效率较传统Faster RCNN算法提高6.7%。在非格式化图像数据快速增加背景下,文章所提方法对于提高电力设备状态图像的识别和处理具有重要意义。
Aiming at the problem that the recognition rate of equipment reduces caused by mismatch between Anchor and target device generated by Faster RCNN in the process of image processing of power equipment,an improved model of Faster RCNN based on connected domain preprocessing was proposed.In order to improve the matching degree between the anchor boxes and the target devices,the connected domain was constructed according to the principle of close pixel value and adjacent locations,and the length width ratio of the connected domain is used as the input information to modify the Anchor area box of RPN(region proposal network,RPN).The experimental results showed that the further matching of recognition area and target equipment area avoided the redundant calculation of anchor frame network.The misrecognition rate of target equipment in this method has been reduced from 2.2%before improvement to 1.9%,with a decrease of 14%.Meanwhile,the calculation efficiency is 6.7%higher than that of the traditional faster RCNN algorithm.Under the background of rapid increase of unformatted image data,the method proposed in this article is of great significance for improving the recognition and processing of power equipment state images.
作者
马静怡
崔昊杨
MA Jingyi;CUI Haoyang(College of Electronics and Information Engineering,Shanghai University of Electric Power,Shanghai 200090,China)
出处
《供用电》
2020年第1期57-61,72,共6页
Distribution & Utilization
基金
上海市地方能力建设项目（15110500900）~~
关键词
电力设备
图像识别
深度学习
连通域
卷积神经网络
power equipment
image recongnition
deep learning
connected domain
convolutional neural network
