第22卷第2期2023年2月Vol.22No.2Feb.2023软件导刊SoftwareGuide基于深度学习的番茄识别与实例分割高倩,诸德宏,封浩(江苏大学电气信息工程学院,江苏镇江212003)摘要:目前番茄采摘主要依靠人工,实现番茄产业机械化和智能化刻不容缓,而番茄检测是最基础也最重要的一步。针对该问题,提出一种基于改进MaskRCNN的番茄检测算法。该算法选择ResNet50和FPN作为主干网络,提出一种新型RoI提取器,并在算法模型中使用空洞卷积(Atrous)。通过Labelme自制番茄数据集,将改进算法在自制数据集上进行训练和测试。结果表明,与FasterRCNN和MaskRCNN模型相比,改进后的模型AP值分别提高了5.5%和4.7%,AR值分别提升了6.8%和4.6%。该算法不仅提高了番茄的识别准确率,还更好地实现了实例分割。关键词:番茄检测;MaskRCNN;新型RoI提取器;空洞卷积;实例分割DOI:10.11907/rjdk.221276开放科学(资源服务)标识码(OSID):中图分类号:TP391.4文献标识码:A文章编号:1672-7800(2023)002-0075-06TomatoRecognitionandInstanceSegmentationBasedonDeepLearningGAOQian,ZHUDe-hong,FENGHao(SchoolofElectricalandInformationEngineering,JiangsuUniversity,Zhenjiang212003,China)Abstract:Atpresent,tomatopickingmainlyreliesonmanuallabor,soitisurgenttorealizethemechanizationandintelligenceofthetomatoindustry,andtomatodetectionisthemostbasicandmostimportantstep.Inresponsetothis,proposeatomatodetectionalgorithmbasedonimprovedMaskRCNN.ThealgorithmselectsResNet50andFPNasthebackbonenetwork,proposesanovelRoIextractor,andusesatrousconvolution(Atrous)inthealgorithmmodel.ThroughtheLabelmeself-madetomatodataset,theimprovedalgorithmwillbetrainedandtest⁃edontheself-madedataset.ComparedwiththeFasterRCNNandMaskRCNNmodels,theimprovedmodelalsoincreasestheAPvalueby5.5%and4.7%,respectively,andtheARvaluethat'sanincreaseof6.8%and4.6%,respectively.Theresultsshowthatitnotonlyimprovestherecognitionaccuracyoftomatoes,butalsobetterachievesinstancesegmentation.KeyWords:tomatodetection;MaskRCNN;newRoIextractor;Atrous;instancesegmentation0引言在我国,随着机器视觉技术的快速发展,图像处理已成为贯穿农业产业链各个阶段的重要技术之一,在选种适配、生长过程、采摘方式及...