摘要
针对变电站巡检机器人在室外进行指针式仪表读数识别时存在检测精度低和读数误差大等问题,提出了一种基于CenterNet和DeepLabv3+的指针式仪表读数识别方法。ECA-Net是一种不降维的局部跨信道交互策略和自适应选择一维卷积核大小的方法,在CenterNet的主干网络引入ECA-Net轻量级注意力机制模块,加强了不同通道之间的特征联系;在DeepLabv3+的ASPP模块并行连接DAMM双注意力机制模块,DAMM模块中的位置注意力模块能够有效模拟出图像位置间的长期上下文依赖信息,将不同局部特征信息连贯起来,提高了语义分割能力;DAMM模块中的通道注意力模块利用不同通道的相关类别特征间的关联性进行不同类别特征强化,提升像素分类精度;利用基于线性变换理论的椭圆透视变换和仿射变换来矫正畸变仪表图像,获取仪表正立图像,提高指针直线拟合角度的精确度,从而减小读数误差。使用该方法进行了大量的仿真与现场测试,结果表明,在仪表检测阶段,所设计模型的mAP比原始模型提高了7.51%;在仪表读数识别阶段,矫正前仪表读数预测值和仪表真实值之间的标称误差为6.0%,平均误差为4.2%,矫正后仪表读数预测值和仪表真实值之间的标称误差为2.0%,平均误差为1.3%,从而验证了所提方法的有效性。
Aiming at the problems of low detection accuracy and large reading error when the substation inspection robot conducts pointer instrument reading recognition outdoors,a pointer instrument reading recognition method based on CenterNet and DeepLabv3+is proposed.ECA-Net is a local cross-channel interaction strategy without dimensionality reduction and adaptive selection of one-dimensional convolutional kernel size method.In the backbone network of CenterNet,ECA-Net lightweight attention mechanism module is introduced,which strengthens the characteristic connection between different channels.In the ASPP module of DeepLabv3+,the DAMM dual attention mechanism module is connected in parallel,and the positional attention module in the DAMM module can effectively simulate the long-term context-dependent information between image positions,connect different local feature information and improve the ability of semantic segmentation.The channel attention module in the DAMM module uses the correlation between the related category features of different channels to strengthen the characteristics of different categories and improve the accuracy of pixel classification.The elliptic perspective transformation and affine transformation based on linear transformation theory are used to correct the distortion of the instrument image,obtain the upright image of the instrument,improve the accuracy of pointer straight line fitting angle,and thus reduce the reading error.A large number of simulations and field tests were carried out using this method.The results showed that in the instrument detection stage,the mAP of the proposed model was increased by 7.51%higher than the original model.In the stage of instrument reading recognition,the nominal error between the predicted value of the instrument reading and the true value of the instrument before correction is 6.0%,and the average error is 4.2%,and the nominal error between the predicted value of the instrument reading and the true value of the instrument is 2.0%and the average error i
作者
黄思远
樊绍胜
王子扬
HUANG Si-yuan;FAN Shao-sheng;WANG Zi-yang(School of Electrical and Information Engineering,Changsha University of Science and Technology,Changsha 410114,China)
出处
《电力学报》
2022年第3期232-243,共12页
Journal of Electric Power
