| 引用本文: | 张森,万吉林,王慧芳,管敏渊,杨斌,李凡.基于注意力机制的卷积神经网络指针式仪表图像读数识别方法[J].电力自动化设备,2022,42(4): |
| ZHANG Sen,WAN Jilin,WANG Huifang,GUAN Minyuan,YANG Bin,LI Fan.Convolutional neural network based on attention mechanism for reading recognition of pointer-type meter images[J].Electric Power Automation Equipment,2022,42(4): |
|
| 摘要: |
| 目前关于指针式仪表图像读数识别的研究大多建立在指针线段检测的基础上，然而该方法流程较多、读数识别效率低。并且仪表图像校准、指针线段拟合等中间过程积累的误差容易使指针倾角偏离真实值。因此从另一角度对基于图像特征映射仪表读数的方法进行了研究，该方法的优势是流程短、效率高。首先构建了融合卷积注意力模块的双路异构卷积神经网络，强化了对仪表图像特征的提取，改善了特征的类型和分布，提高了仪表读数识别的准确率；然后采取了软区间分级回归的策略，极大地简化了模型的体积，使得模型易于部署；最后通过算例对比了所提方法和基于指针线段检测的深度学习、机器学习模型识别仪表读数的准确率和效率。算例表明，所提方法在仪表图像读数识别准确率和效率之间取得了较好的平衡。 |
| 关键词: 指针式仪表 读数识别 卷积神经网络 注意力机制 分级回归 软区间 |
| DOI：10.16081/j.epae.202112027 |
| 分类号:TM63;TM930 |
| 基金项目:国网湖州供电公司科技项目(2019-HUZJTKJ-09) |
|
| Convolutional neural network based on attention mechanism for reading recognition of pointer-type meter images |
|
ZHANG Sen1, WAN Jilin1, WANG Huifang1, GUAN Minyuan2, YANG Bin2,3, LI Fan2
|
|
1.College of Electrical Engineering, Zhejiang University, Hangzhou 310027, China;2.Huzhou Power Supply Company of State Grid Zhejiang Electric Power Co.,Ltd.,Huzhou 313000, China;3.Huzhou Electric Power Design Institute Co.,Ltd.,Huzhou 313000, China
|
| Abstract: |
| At present, most researches on the reading recognition of pointer-type meters are based on the detection of pointer line segments. However, this method has many processes and low reading recognition efficiency. Besides, the errors accumulated in the intermediate processes such as meter image calibration and pointer segment fitting can easily cause the pointer inclination angle to deviate from its true value. The image feature mapping method for meter readings is studied from another perspective because this method has the advantage of fewer processes and higher efficiency. Firstly, a two-stream heterogeneous convolutional neural network fused with CBAM(Convolutional Block Attention Module) is constructed, which strengthens the extraction of meter image features, improves the type and distribution of features and the accuracy of meter reading recognition. Then a soft stagewise regression strategy is adopted, which greatly simplifies the size of the model and makes the model easy to deploy. Finally, the proposed method is compared with deep learning models based on the pointer line segment detection and other machine learning models on the recognition accuracy and efficiency of pointer-type meter reading recognition by example. Results show that the proposed method maintains a good balance between the accuracy and efficiency of the meter image reading recognition. |
| Key words: pointer-type meter reading recognition convolutional neural network attention mechanism stagewise regression soft range |
