基于改进的EfficientDet的布匹疵点识别杨连贺,张超(天津工业大学计算机科学与技术学院,天津300387)摘要:为了准确而高效地识别出布匹各种疵点的种类,采用改进的EfficientDet算法进行布匹疵点识别。首先采取改进的Ostu阈值分割算法进行特征边缘的检测,采用非极大值抑制方法对边缘进行筛选,确定候选区域;然后采用筛选器对候选区域的疵点进行识别和分类,其中筛选器采用改进的EfficientDet算法。改进的EfficientDet算法与其他优秀的目标检测算法以及原算法进行了比较。结果表明,改进的Ostu分割算法相较于传统算法不仅可以在更多的布匹图像中更准确地识别疵点区域,而且抑制了假边缘现象;该模型规模是几种算法中最小的,识别准确率达到94%,高于目前最优算法4个百分点。关键词:疵点检测;迁移学习;Ostu;目标检测;EfficientDet中图分类号:TS101.97;TP391.41文献标志码:A文章编号:员远苑员原园圆源载(圆园23)园4原园园71原06FabricdefectrecognitionbasedonimprovedEfficientDetnetworkYANGLianhe,ZHANGChao(SchoolofComputerScienceandTechnology,TiangongUniversity,Tianjin300387,China)Abstract:Inordertoidentifyallkindsoffabricdefectsaccuratelyandefficiently袁theimprovedEfficientDetalgorithmisusedforfabricdefectsrecognition.TheimprovedOstuthresholdsegmentationalgorithmisadoptedtodetectfea鄄tureedges袁andnon-maximumsuppressionisusedtoscreenedgestodeterminecandidateregions.Thenafilterisusedtoidentifyandclassifythedefectinthecandidatearea.ThefilterusesanimprovedEfficientDetalgorithm.TheimprovedEfficientDetalgorithmiscomparedwithotherexcellenttargetdetectionalgorithmsandtheoriginalalgorithm.TheexperimentalresultsshowthattheimprovedOstusegmentationalgorithmcanaccuratelyidentifythedefectareasinmorefabricimagesandsuppressthefalseedgephenomenoncomparedwiththetraditionalal鄄gorithm.Themodelsizeisthesmallestamongthecentralizedalgorithms袁andtherecognitionaccuracywas94%袁beatingthebestalgorithmby4percentagepoints.Keywords:defectdetection曰transferlearning曰Ostu曰targetdetection曰EfficientDet虽然我国的纺织品制造与出口量位居世界前列,但是布匹疵点的识别主要还是依赖于人工。检测工不可能长时间专注并且高效地工作,因而不能很好地把控布匹的...