面向豆类作物病害的改进YOL0v10检测算法

打开文本图片集

中图分类号：S435.2 文献标识码：A 文章编号：2095-5553（2025）10-0203-07

Abstract：Toenablereal-timeandacuratedetectionof bean diseases，an improvedYOLOvlOobjectdetectionalgorithmwas developed.First，the DCNv2 module was introduced to replace the conventional CBS module in the neck network，which enhanced themodel’s abilitytodetect diseased crops with varying shapes andspatial distributions.Subsequently，a BiFPN structure wasincorporatedtostrengthenthemodelscapacityfordistinguishing featuresacross diferent targetscalesbyemploying a multi-scale feature weighting and fusion strategy.Additionaly，a multi-scale fusion detection head basedonan attntion mechanism was designedtomakefulluseofthemultiscaleinputfeatures provided bytheneck network，enabling moreffective detectioofobjects at varying sizes.Experiments wereconductedusingapubliclyavailable bean diseasedataset.Theimproved model achieved an mAP@0.5 of 72.1% and an mAP@0.5：0.95 of 47.1% ，reflecting gains of 3% and 4.9% ，respectively， compared to the original YOLOv1O model，without increasing computational costs.Overall，the improved algorithm can significantlyenhancetheaccuracyofbean diseasedetectionandprovidevaluablesupportfortheeffectivecultivationand large-scale promotion of bean crops.

Keywords：bean diseases；object detection；deformable convolution；multi-scale features；attention mechanism

0 引言

豆类植物作为一种分布广泛的农业作物，在全球农业中起到重要的基础性作用[1]。（剩余11893字）