河北工业大学学报JOURNALOFHEBEIUNIVERSITYOFTECHNOLOGY2023年2月February2023第52卷第1期Vol.52No.1基于残差网络与特征融合的改进YOLO目标检测算法研究王彤，李琦（河北工业大学电子信息工程学院，天津300401）摘要以深度学习为基础的YOLO目标检测技术因检测速度快，而广泛应用于实时目标检测领域中，但其检测准确率不高，尤其是对小物体的检测能力较差。针对上述问题，本文提出一种改进模型——R-YO⁃LO。该模型将残差单元引入YOLO目标检测，既可以通过增加网络的深度，提高网络的准确性，又可以利用残差网络的快捷连接方式，以保证检测的实时性。同时结合CBNet结构，增强语义信息，进一步提高R-YOLO的准确性。最后在改进的YOLO模型中通过特征金字塔融合，结合不同阶段卷积层输出的特征信息，使得融合后的特征图同时具有深层次的语义信息和浅层次的位置信息，以提高对小物体的检测准确性。在Pascal数据集上的实验显示R-YOLO在准确率上较YOLO提高了7.6个百分点，对小物体的检测结果更准确。结果表明，残差单元和特征金字塔融合的引入有效改进了YOLO网络模型的检测性能。关键词深度学习；目标检测；YOLO；残差网络；特征融合；CBNet中图分类号TP319.4文献标志码AResearchonimprovedYOLOtargetdetectionalgorithmbasedonresidualnetworkandfeaturefusionWANGTong,LIQi（SchoolofElectronicsandInformationEngineering,HebeiUniversityofTechnology,Tianjin300401,China）AbstractTheYOLOtargetdetectiontechnologybasedondeeplearningiswidelyusedinthefieldofreal-timetargetdetectionwithitsfastdetectionspeed,butitsdetectionaccuracyisnothigh,especiallyforsmallobjects.Inresponsetotheaboveproblems,thispaperproposesanimprovedmodel-R-YOLO.ThemodelintroducestheresidualunitintoYOLOtargetdetection,whichcannotonlyincreasethedepthofthenetworktoimprovetheaccuracyofthenetwork,butalsousethefastconnectionmethodoftheresidualnetworktoensurethereal-timedetection.CombinedwiththeCBNetstruc⁃ture,thesemanticinformationisenhancedandtheaccuracyofR-YOLOisfurtherimproved.Finally,throughfeaturepyr⁃amidfusionintheYOLOmodel,combinedwiththefeatureinformationoutputbytheconvolutionallayersatdifferentstages,thefusedfeaturemaphasbothdeepsemanticinformationandshallowlocationinformationtoimprovedetectionaccuracy...
