基于改进DCGAN的棉叶螨为害图像数据增强方法

打开文本图片集

中图分类号：TP391.4 文献标识码：A 文章编号： 1000-4440（2025）05-0916-1

Abstract：Toaddressthe insufficientand imbalancedsample sizes of cottonleaf mitedamage images at different severity levels，reducedatacollectioncosts，andenhancethequalityanddiversityofimagesgeneratedbygenerativeadversarialnetworks，thisstudyproposedanimprovedDCGAN-baseddataaugmentationmethodforcoton leaf mitedamageimages.Basedon theoriginalmodel，ategorylabelswereintroducedtoeableargetedgenerationofimagesfordiferentdamagelevels，ectively resolvingtheissueofclassimbalance.Thetraditional directconnectionstructurewasreplacedwitharesidualstructuretoenhance the model’sabilty to leam complex mapping relationships，avoid gradient vanishing problems，and improvethequalityof generatedimages.Aditionally，theconvolutional blockatentionmodule（CBAM）wasembedded intheconvolutionallayers to strengthenthe model’s capacity to extract key featuresof cottonleafmitedamageimages，further enhancingthequality anddiversityofgenerated images.Lastly，theWasserstein distancewith gradient penaltywas employed as the loss function，avoiding the problem of mode collapse and enhancing thetraining stability ofthe model. The improved DCGAN

modeloutperformedtheoriginalmodelintemsof trainingstabilityandimagequality.Itsgeneratedimagesachievedhigherinceptionscore（，.1）rheticptiodisace（F，5.12），elieptionsane（K，6）dsuallarityindex measure（SSM，O.82）thanthosegeneratedbyotherclassicdataaugmentationmodelsWhentrainingtheDenseNet121 model with the dataset generated by the improved DCGAN model，the average clasification accuracy reached 88.02% ，which washighrthanthatofDenseNet-121modelstrainedwithdatasetsgeneratedbytraditionalaugmentationmethodsandothermodels.Thisstudy provides technical support for intelligent monitoring of agricultural pests and diseases.

Key words：cotton leaf mite；damage degree；deepconvolutional generative adversarial network （DCGAN）；image data augmentation

新疆作为中国最大的优质棉生产基地，棉花产量占全国总产量的 90% 以上，其生产稳定性直接关系到国家棉花战略安全与区域经济发展[1]。（剩余14219字）