基于鉴别注意力融合的仪表细粒度分类方法

来源：计算机技术与发展杂志2023年第07期北京时间：

作者：孙荣艳;李晓明

单位：太原科技大学 计算机科学与技术学院,山西 太原 030024 Author(s): SUN Rong-yan;LI Xiao-ming School of Computer Science and Technology,Taiyuan University of Science and Technology,Taiyuan 030024,China 关键词: 双线性融合;正交损失;类激活热力图;YOLOv5;工业仪表 Keywords: bilinear fusion;orthogonal loss;class activation map;YOLOv5;industrial meter 分类号: TP391. 4 DOI: 10. 3969 / j. issn. 1673-629X. 2023. 07. 025 摘要: 基于视觉的仪表自动巡检读数是一项重要的研究内容;仪表的读数通常通过检测指针的位置来确定;不同规格型号的仪表,指针指向相同位置所代表的具体读数不同,因此,预先识别仪表的详细类别是进行自动读数的重要前提。 为提升仪表自动读数的便捷性和准确率,提出了一种基于鉴别注意力融合的仪表细粒度分类方法。 首先利用 YOLOv5 对仪表表盘进行粗提取,随后使用提出的模型对表盘进行细粒度识别;无需标注显著特征,在特征提取器上添加鉴别注意力模块,补充浅层空间和位置信息对仪表细粒度分类的引导作用;生成的鉴别粒度注意力图和骨干网络最后一层输出的特征图进行双线性融合,生成特征矩阵;引入正交损失,对生成的特征矩阵进行约束处理;构建仪表细粒度分类数据集。 理论分析和实验结果表明,所提仪表细粒度分类方法提高了网络对表盘鉴别粒度区域的识别能力,有效改善了仪表细粒度分类的性能,为后续仪表智能准确的读数提供了保证。 Abstract: Automatic inspection reading of instrument based on vision is an important research content. The automatic reading of theindustrial meter is usually determined by detecting the position of the pointer. For meters of different specifications?and categories,thespecific readings represented by the pointer pointing to the same position are different. Therefore,?it is an important prerequisite forautomatic reading to identify the detailed categories of meters in advance. In order to improve the convenience and accuracy of automaticmeter reading,a fine-grained meters classification method based on discriminant attention fusion is proposed. Firstly,the meter panel isroughly extracted by YOLOv5,and then the proposed model is used for fine-grained recognition of the meter panel. There is no need tomark significant features,and the discriminant attention module is added to the feature extractor to supplement the guiding role of shallowspace and location information on the fine granularity classification of meters. The discriminant granularity attention map generated isbilinear fused with the feature map output from the last layer of the backbone network to generate the feature matrix. Orthogonal loss isintroduced to constrain the generated feature matrix. The meter fine-grained classification data set is built. Theoretical analysis and experimental results show that the proposed meter fine-grained classification method effectively improves the performance of the meter fine-grained classification,and provides a guarantee for the subsequent intelligent and accurate reading of the meter.

摘要：基于视觉的仪表自动巡检读数是一项重要的研究内容;仪表的读数通常通过检测指针的位置来确定;不同规格型号的仪表,指针指向相同位置所代表的具体读数不同,因此,预先识别仪表的详细类别是进行自动读数的重要前提。 为提升仪表自动读数的便捷性和准确率,提出了一种基于鉴别注意力融合的仪表细粒度分类方法。 首先利用 YOLOv5 对仪表表盘进行粗提取,随后使用提出的模型对表盘进行细粒度识别;无需标注显著特征,在特征提取器上添加鉴别注意力模块,补充浅层空间和位置信息对仪表细粒度分类的引导作用;生成的鉴别粒度注意力图和骨干网络最后一层输出的特征图进行双线性融合,生成特征矩阵;引入正交损失,对生成的特征矩阵进行约束处理;构建仪表细粒度分类数据集。 理论分析和实验结果表明,所提仪表细粒度分类方法提高了网络对表盘鉴别粒度区域的识别能力,有效改善了仪表细粒度分类的性能,为后续仪表智能准确的读数提供了保证。

关键词：双线性融合;正交损失;类激活热力图;YOLOv5;工业仪表