引入注意力机制的偏转字符识别算法  

Deflection character recognition algorithm introducing attention mechanism

作　　者：王仁睿 张宝龙[1] 李丹[1] 马煜峰 张鑫 乔高学 张志强 WANG Renrui;ZHANG Baolong;LI Dan;MA Yufeng;ZHANG Xin;QIAO Gaoxue;ZHANG Zhiqiang(School of Electronics and Automation,Tianjin University of Science and Technology,Tianjin 300222,China;Tianjin Elco Automation Co.Ltd.,Tianjin 300385,China)

机构地区：[1]天津科技大学电子信息与自动化学院,天津300222 [2]天津宜科自动化股份有限公司,天津300385

出　　处：《液晶与显示》2024年第10期1322-1331,共10页Chinese Journal of Liquid Crystals and Displays

摘　　要：由于传统人工抄表成本高且受恶劣环境限制,使得智能抄表成为以后的发展方向。常见的电表主要可分为机械字轮电表和液晶电表两类。其中,由于机械字轮电表存在的偏转问题,导致其在字符识别的过程中出现了字符特征信息缺失的情况,从而使该表型识别准确率较低。为了解决此问题,本文对YOLOv5识别算法的骨干网部分进行了修改,提升了算法对机械字轮电表偏转字符的识别效果。首先,在网络模型中引入了CBAM注意力机制,提升了网络模型对偏转字符的特征提取能力。其次,将切片操作(Focus)替换成一个6×6的卷积,并使用更快的SPPF池化结构替换了原先的SPP池化结构以提升算法的运算速度。为了测试模型的识别效果,采集了329张电表偏转字符样本进行实验,整体识别准确率可以达到99.4%。同时采集了1500张液晶电表样本对模型的泛化性进行测试,识别准确率达到了99.6%。实验结果表明,本文方法解决了偏转字符识别率低的问题,同时验证了识别模型具有很强的泛化性。Due to the high cost and harsh environmental limitations of traditional manual meter reading,intelligent meter reading has become the future development direction.The common types of electric meters can be mainly divided into mechanical word-wheel and liquid crystal electric meters.Among them,due to the deflection problem of mechanical word-wheel electric meters,there is a lack of character feature information in the process of character recognition,which leads to a low accuracy of this electricity meter type recognition.In order to solve this problem,this paper modifies the backbone network of YOLOv5 recognition algorithm,which improves the recognition effect of the algorithm on deflection characters of mechanical word-wheel electricity meter.Firstly,CBAM attention mechanism is introduced into the network model,which improves the feature extraction ability of the network model for deflected characters.Secondly,the Focus operation is replaced by a 6×6 convolution,and the original SPP pooling structure is replaced by a faster SPPF pooling structure to improve the operation speed of the algorithm.In order to test the recognition effect of the model,329 deflection character samples of electric meters are collected for experiments,and the overall recognition accuracy can reach 99.4%.At the same time,1500 samples of liquid crystal electric meters are collected to test the generalization of the model,and the recognition accuracy reaches 99.6%.The experimental results show that this method solves the problem of low recognition rate of deflected characters,and verifies that the recognition model has strong generalization.

关 键 词：电表读数识别 YOLOv5 CBAM注意力机制 池化结构 偏转字符 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]