基于改进YOLOv8n的轻量化水稻病害检测方法

打开文本图片集

中图分类号：S435.111；TP391.4 文献标识码：A 文章编号：2095-5553（2025）12-0094-09

Abstract：To address the problemsof high model complexity，significant scalediferences in diseasefeatures，and low detectionacuracyinexisting rice disease detection models，animproved lightweight rice diseasedetectionalgorithm YOLOv8n—ADMW is proposed.First，the ADown lightweight downsampling module is introduced to replace the CBS modulein thenetwork，which helps themodelcapture disease featuresandreducecomputationalload，achieving model lightweightig.Then，thedynamicupsamplingoperatorDysamplebasedonpointsampling isusedtoreplacetraditional upsampling methods，adaptivelylearningsampling parameters accordingto inputfeaturecontenttoreduce thelossof disease edge and feature information.The C2f—MSBlock module is designed toreplace the C2f module inthe neck network， furtherenhancing multiscalediseasefeature fusioncapabilityand improvingdetectionacuracyformulti-scaletargets.The original CIoU loss functionof YOLOv8n is changed to thedynamic non-monotonic focusing mechanism WIoUv3，which improvesthemodel’s precise localizationcapabilitythrough gradient gain alocation strategy.Experimental results showthat the improved lightweight model achieves mAP@0.5 and of 91.4% and 56.8% respectively，representing improvements of 4.5% and 3.5% compared to the YOLOv8n baseline model. The model parameters and size are only 2.4M and5MB respectively，representing reductionsof 20% and 16.7% compared to before improvement.The YOLOv8n—ADMW algorithm maintains model lightweighting while improving detectionaccuracy，enabling effective detection of rice diseases and providing technical support for subsequent diagnosis of diferent rice diseases.

Keywords：rice disease；target detection； lightweight； loss function；multi-scale feature fusion

0 引言

水稻是全球范围内最重要的粮食作物之一，水稻的产量直接影响到粮食安全。（剩余13799字）