2023年第5期仪表技术与传感器InstrumentTechniqueandSensor2023No．5收稿日期:2022－07－19基于双动态头SparseＲ-CNN的表面缺陷检测算法郑亚睿，蒋三新(上海电力大学电子与信息工程学院，上海201306)摘要:为了减少缺陷检测中的冗余检测，提出基于双动态头SparseＲ－CNN的缺陷检测算法，2个动态头的责任不同:第1个负责不同尺度和空间的特征提取，第2个负责匹配可学习的提议特征。为了更好地提取图像细节信息，改进特征金字塔(FPN)为特征金字塔网格(FPG)，并且与第1个动态头相结合进行特征提取。其次，提出了交流注意力来改进检测阶段的多头自注意力模块，减少随着迭代注意力图相似导致建模能力下降的问题。最后，改进边框回归损失函数GIoU为Alpha－CIoU，加速收敛并提升检测的精度。实验结果表明:算法在晶圆和热轧钢2种表面缺陷数据集上都取得很好效果，平均精度分别为94．3%和88．1%。关键词:表面缺陷检测;动态头;稀疏预测;注意力机制;标签匹配;端到端预测中图分类号:TP391;TN407文献标识码:A文章编号:1002－1841(2023)05－0097－09SurfaceDefectDetectionAlgorithmBasedonDualDynamicHeadSparseＲ－CNNZHENGYa-rui，JIANGSan-xin(SchoolofElectronicandInformationEngineering，ShanghaiElectricPowerUniversity，Shanghai201306，China)Abstract:AdefectdetectionalgorithmbasedondualdynamicheadSparseＲ－CNNwasproposedtoreduceredundantdetec-tionindefectdetection．Theresponsibilitiesofthetwodynamicheadsweredifferent，thefirstwasresponsibleforfeatureextractionindifferentscalesandspaces，andthesecondwasresponsibleformatchinglearnableproposedfeatures．Thefeaturepyramidnet-work(FPN)wasimprovedtothefeaturepyramidgrid(FPG)andcombinedwiththefirstdynamicheadforfeatureextraction．Secondly，speaking-headattention(SHA)wasproposedtoimprovethemulti-headself-attentionmoduleinthedetectionstagetoreducethedeclineofmodelingabilitywiththesimilarityoftheiterativeattentiongraph．Finally，thelossfunctionGIoUofboun-dingboxregressionwasimprovedbyAlpha-CIoU，whichacceleratedconvergenceandimprovestheaccuracyofdetection．Theex-perimentalresultsshowthatthealgorithmhasachievedgreatresultsonthesurfacedefectdatasetsofthewaferandhotrolledsteel，andtheaverageaccuracyis94．3%and88．1%respectivel...
