基于Transformer的多尺度工件编码识别算法  

Multi-scale recognition algorithm for workpiece code labels based on Transformer structure

作　　者：熊新炎 马宏伟 张良 XIONG Xinyan;MA Hongwei;ZHANG Liang(College of Light Industry,Harbin University of Commerce,Harbin 150028,China)

机构地区：[1]哈尔滨商业大学轻工学院,哈尔滨150028

出　　处：《哈尔滨商业大学学报（自然科学版）》2024年第5期536-543,共8页Journal of Harbin University of Commerce：Natural Sciences Edition

摘　　要：提出了一种基于Transformer结构的多尺度工件编码识别算法,旨在解决当前工件编码识别中存在的挑战与局限性.描述了多尺度特征提取和融合的实现方法,重点介绍了Transformer模块的优化策略.通过卷积和池化操作获得了一组在不同尺度和层级上的特征,针对这些特征引入了一个创新的缩放因子,用于调整Transformer模块中的注意力权重,更精确地捕捉和融合不同尺度的特征信息.提出了缩放因子计算方法,该方法直接依赖于查询(Query)和键(Key)的信息,可以更为直观地反映出不同尺度特征在注意力计算中的重要性.实验结果表明,此方法在处理多尺度工件编码特征时表现出了较高的准确率和稳健性,可有效提升工件编码识别的性能.This paper proposed a multi-scale workpiece encoding recognition algorithm based on the Transformer structure,aiming to address the existing challenges and limitations in current workpiece encoding recognition.Provided an overview of the application of machine vision in workpiece encoding recognition and relevant knowledge regarding the Transformer structure.By deeply exploring the roles of multi-scale features and Transformer modules in workpiece encoding recognition,this paper details the implementation methods for multi-scale feature extraction and fusion,and prominently introduced the optimization strategies for the Transformer module.This paper obtained a set of features at different scales and levels through convolution and pooling operations.Subsequently,an innovative scaling factor was introduced to adjust the attention weights in the Transformer module to more accurately capture and fuse features of different scales.And proposed a new method for calculating the scaling factor,which directly depends on the information of Query and Key,and can more intuitively reflect the importance of different scale features in attention computation.Additionally,the results of which demonstrate that our method exhibits higher accuracy and robustness when dealing with multi-scale workpiece encoding features,effectively enhancing the performance of workpiece encoding recognition.

关 键 词：工件编码 多尺度特征 缩放因子 注意力权重 特征融合 TRANSFORMER 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]