改进HED网络的金属零件二维条码分割方法  

Segmentation method of DataMatrix codes marked in metal parts based on improved HED network

作　　者：李建伟[1] 吕娜 郭宏[2] 刘成波 LI Jian-wei;LYU Na;GUO Hong;LIU Cheng-bo(School of Computer Science and Technology,Taiyuan University of Science and Technology,Taiyuan 030024,China;School of Mechanical Engineering,Taiyuan University of Science and Technology,Taiyuan 030024,China)

机构地区：[1]太原科技大学计算机科学与技术学院,山西太原030024 [2]太原科技大学机械工程学院,山西太原030024

出　　处：《计算机工程与设计》2023年第3期859-865,共7页Computer Engineering and Design

基　　金：山西省回国留学人员基金项目(HGKY2019079)。

摘　　要：针对复杂环境金属件标刻的DM码因磨损、腐蚀等原因,导致传统图像分割算法难以精确定位的问题,提出一种基于改进HED网络的金属零件二维条码分割方法。在原HED网络上对主干特征提取网络改进,采用空洞卷积扩大感受野保留DM码全局信息;改变特征融合模块,增加两层卷积运算充分融合深监督模块输出的多尺度特征,提取区域轮廓完成DM码区域分割;使用LSD算法寻找区域分割图中的直线,经过直线聚类减少非感兴趣直线产生的干扰,实现二维条码区域精确分割定位。实验结果表明,该模型在全局最佳(ODS)和单图最佳(OIS)精度评定中F1值分别达到0.813和0.825,平均定位准确率达到97%。Aiming at the difficult localization and inaccurate image segmentation issues of DataMatrix code marked on metal parts in complex background resulted from wear and corrosion etc.,an accurate segmentation method of DataMatrix codes based on an improved HED network was proposed.Improvements were made on the basis of the HED network structure,the backbone feature extraction network was improved,the dilated convolution structure was used to expand the perceptual field and preserve the global information of the DM code.The model feature fusion module was changed to add two-layer convolution operations to fully fuse the multi-scale feature output of the deep supervised module,the improved model was used to extract the contour of the area to complete the DM code area segmentation.The LSD algorithm was used to find all the lines in the area segmentation map,the impact of non-interesting lines on the location was reduced through linear clustering.Experimental results show that this model gets the highest F1-values of 0.813 and 0.825 in ODS and OIS,respectively.The average localization accuracy can reach up to 97%.

关 键 词：DataMatrix码 HED网络模型 空洞卷积 特征融合 多尺度特征提取 LSD算法 图像分割 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]