基于改进DenseNet的棉叶螨危害等级识别研究

打开文本图片集

中图分类号：S24；TP391.4 文献标识码：A 文章编号：2095-5553（2026）02-0202-08

Abstract：Inresponseto the labor-intensiveand time-consuming nature，aswell as the lagging issues inherentin traditional manual diagnosisandgradingmethodsforcottonleaf mites，anovelcotonleafmitehazardlevelrecognition modelbasedonanimprovedDenseNet—121isproposed.Followingthestandardclasificationcriteriaforcotonleaf mite damage，images of coton leaves with varying hazard levels under both uniform and natural backgrounds were collected. Thedataset wasaugmented to simulate the influenceof diverseweather conditions，shoting angles，and device noise during imageacquisition.Considering thehigh similarity betwendiferent mite damagelevelsandtheresulting identification chalenges，we implemented three keyimprovements to the DenseNet—121 model.First，the7 ×7 （204 convolution kernel intheinitialconvolution layerwasreplacedwithan Inceptionmoduletoenhancethefeatureextraction capabilityof theshallownetworklayers.Second，a SimAMatentionmechanismwas introducedafter the Transition Layerto emphasize cotton leaf mite hazard features and suppressbackground features.Lastly，DropBlock regularization was appliedafter the DenseLayertoenhance the model’srobustness andpreventoverfiting.Theresultsdemonstrate that the original model achieves a recognition accuracy of 90.76% on the augmented dataset，representing an improvement of 4.21 percentagepoints over the original model. Individualy，data augmentation and the threeimprovement strategies increase the recognition accuracyof the model by1.47，2.74，2.37and1.86 percentagepoints，respectively.The comprehensive performance of the model is notably superior to that of other models such as VGG16 and ResNetE Keywords：cotton leaf;mite hazard level; improved DenseNet—l21；attention mechanism；regularization

0 引言

棉花是我国最重要的农作物之一，新疆是我国最大的棉花生产基地[1.2]。（剩余12544字）