改进YOLOv4框架的胃息肉检测①吴宇杰,肖满生,范明凯,胡一凡(湖南工业大学计算机学院,株洲412007)通信作者:肖满生,E-mail:xiaomansheng@hut.edu.cn摘要:在计算机视觉的内窥胃部息肉检测中,高效提取小型息肉图像特征是设计深度学习的计算机视觉模型一个难点.针对该问题,提出了一种YOLOv4改进的YOLOv4-polyp检测模型.首先在YOLOv4的基础上,引入CBAM卷积注意力模块增强模型在复杂环境的特征提取能力;其次设计出轻量级CSPDarknet-49网络模型,在降低模型复杂度的同时提高检测精度和检测速度;最后根据胃息肉数据集的特点,采用K-means++聚类算法对胃息肉数据集进行聚类分析,得到优化后的锚框.实验对比结果表明,YOLOv4-polyp对于经典YOLOv4模型在保持检测速率不变的同时,在两个数据集中平均检测精度分别提升了5.21%和2.05%,表现出良好的检测性能.关键词:YOLOv4;注意力机制;K-means++;目标检测引用格式:吴宇杰,肖满生,范明凯,胡一凡.改进YOLOv4框架的胃息肉检测.计算机系统应用,2023,32(2):250–257.http://www.c-s-a.org.cn/1003-3254/8931.htmlImprovedYOLOv4FrameworkforGastricPolypDetectionWUYu-Jie,XIAOMan-Sheng,FANMing-Kai,HUYi-Fan(SchoolofComputerScience,HunanUniversityofTechnology,Zhuzhou412007,China)Abstract:Inendoscopicgastricpolypdetectionbasedoncomputervision,efficientlyextractingthefeaturesofimagesofsmallpolypsisadifficultyinthedesignofadeeplearning-basedcomputervisionmodel.Tosolvethisproblem,thisstudyproposesadetectionmodelbasedonanimprovedyouonlylookonceversion4(YOLOv4),namelyYOLOv4-polyp.Specifically,onthebasisofYOLOv4,thisstudyaddsaconvolutionalblockattentionmodule(CBAM)toenhancethefeatureextractioncapabilityofthemodelincomplexenvironments.Then,alightweightCSPDarknet-49networkmodelisdesignedtobothreducethecomplexityofthemodelandimproveitsdetectionaccuracyanddetectionspeed.Finally,consideringthecharacteristicsofthegastricpolypdatasets,theK-means++clusteringalgorithmisusedfortheclusteranalysisofthegastricpolypdatasetsandtheattainmentoftheoptimizedanchorbox.TheexperimentalcomparisonresultsshowthatcomparedwiththeclassicalYOLOv4model,theproposedYOLOv4-polypachievesfavorabledetectionperformanceonthetwodatasetsasitimprovestheaveragedetectionaccu...
