基于Swin-PIDNet的纸质工程制图线型识别方法

打开文本图片集

关键词：PIDNet；SwinTransformer；线型识别；纸质工程制图；迁移学习；混合空洞卷积中图分类号：TP391.41 文献标志码：A 文章编号：1001-3695（2026）01-038-0313-08doi： 10.19734/j. issn.1001-3695.2025.04.0145

Line pattern recognition method of paper engineering drawing based on Swin-PIDNet

Zhu Wenbo†，Chen Longfei，Chi Yulun （SchoolofMechanicalEngineering，UniversityofShanghaifor Scienceand Technology，Shanghai 2ooo93，China）

Abstract：Linepaternrecognitionpresentstheprimarychalengeinidentifying imagesof paperengineeringdrawings.Addresing issuessuchaslowline-type standardization，longspan，andsmallrelative sizecomparedtothebackground，this paper proposeda Swin-PIDNetmodelfor linepatternrecognitionof paper engineeringdrawing.Themodelreplacedtheoriginal PIDNet backbone network with Swin Transformer，reducing downsampling whileenhancing the model’slong-ranging modeling capability.Thispaperproposedastage-by-stageunfreeingtransferlearing toimprovethetraining eficiencyandaccuracyof the modelforlinepaternrecognitionandtosmooththemodel training processTohandletheslendercharacteristicsof engineeringdrawing lines，itembeddedtheattentionmoduleEMAintothePAHDCmodule，mitigatingtheproblemofbackground informationoverwhelming linefeature information．Furthermore，toaddress class imbalance inlinepatern，itconstructeda combinedloss function integrating weightedFocallossand Dice lossfortraining Swin-PIDNet.Experimentalresultsdemonstrate that the proposed model achieves an MIoU of 87.02% ，MPA of 95.42% ，and F1 -score of 96.57%.Compared toother models，Swin-PIDNetexhibitssuperiorlinepaternrecognitioncapability，holdingsignificanttheoreticalandpracticalvalue for paper engineering drawing image analysis.

Key words：PIDNet；Swin Transformer；line patternrecognition；paper enginering drawing；transfer learning；hybrid dilasted convolution

0 引言

在工程图纸中，线型是传递信息、表达设计意图的重要手段。（剩余17523字）