基于改进YOLOv11的焊缝缺陷高精度检测方法

打开文本图片集

关键词：YOLOvl1；焊缝缺陷检测；鲁棒纹理差异归一化；边缘轮廓感知；频域特征调制中图分类号：TP391.4 文献标志码：A 文章编号：1001-3695（2025）10-038-3192-09doi：10.19734/j. issn.1001-3695.2025.01.0029

High-precision weld defect detection based on improved YOLOv11

Wang Xiaoting’，Liu Tingting2，3+ （1.SchooloffoatioEnin，ifengUesitKifegHan5O，in;2olofmpuer&ftae， SIASUniersityZengo4in；3HnnItellgenMfacturin&igitalTinEgininReseacCete， ，China）

Abstract：Aimingattheproblemsoflowimagecontrast，complexdefectmorphology，defectdiversity，and imbalancedpositiveandnegativesampleratioinwelddefectdetection，thispaperproposedanimprovedhigh-precisionwelddefectdetection methodbasedontheYOLOv11network framework.Firstly，the method improvedtheperceptionabilityoftexturedifrences inlowcontrastareasof weldimages byintroducingarobust texture differencenormalization mechanism，andenhancedthedetection accuracyof small defects.Then，the method used a boundarycontour awarenessdecoder moduletooptimizethe boundaryinformationofthedetectionbox，aiming toimprovetheacuracyoflocatingcomplexdefects.Subsequently，themethod utilizedthedesignedfrequency-domainfeaturemodulationstrategytofusethespatialandfrequencydomainfeatures，enhancingthe perceptualabilityandsensitivityfordetectingsubtle defects in welds.Finally，thispaperproposedadynamicbalance loss function forpositiveand negativesamples tooptimize the problemof imbalanced sample distribution during the training process ydynamicalladjusting sample weights.The experimentalresultsshowthat the proposed methodhas higherdetection accuracyandstrongerrobustnesscompared totheoriginal YOLOv11model，improvingtheaccuracyrecall，andmAPindicatorsby2.7，3.4，and2.6percentpoints，respectively，andstillmaintainsahighdetectionspeedof43.7fps，whichcanbetter adapt to the detection of weld defects in complex industrial environments.

Keywords：YOLOv11；welddefectdetection;robust texturediference normalization；boundarycontourawareness；frequency domain feature modulation

0 引言

焊接是通过加热或者加压的方式将两个或多个金属部件熔接在一起的过程，广泛应用于航空、航天、汽车、船舶、建筑、能源和机械制造等工业领域[1～3]。（剩余22439字）