基于改进CenterNet的农作物害虫无锚检测算法研究

打开文本图片集

中图分类号：TP391.4；S5 文献标识码：A 文章编号：2095-5553（2025）10-0183-09

Abstract：Pest management inagricultural production playsa keyrole in ensuringcrop yieldandquality.Although pest detectiontechnologybasedonConvolutionalNeuralNetworkhasmadesomeprogressinrecentyears，itstllfaces two majorchallenges.First，somepestspeciesarehighlysimilarinappearance，whichmakes itdificult toclasifyanddentify. Secondly，themulti-scalecharacteristicsofpestobectsleadtoalargenumberoffalsenegativedetections，especiall nthe identificationof smallpests.To solve these problems，areal-time，anchorfreeimproved CenterNet pest detection model was proposed inthis study.TheCBAMatention mechanism wasembedded inthebackbone network toefectivelyimprove the clasification accuracy offeatures.At the same time，a multiscale feature fusion module named MFF was aded to the neck network toachieveeficient integrationoffeaturemapsofdiferentscales.Alargenumberofexperimentaldatashows that the mAP of the model on the Baidu AI insect dataset and the highly challenging IPlO2 dataset is as high as 98.6% and （204号 89.7% respectively，and the inference speed on the two datasets is more than3O frames/s ，showing remarkable real-time performance.Compared with the existing mainstream methods，the improved CenterNet pest detectionmodelshows outstandingadvantages intheaccurateidentificationof pests incomplexagricultural environments，andhasawiderangeof application potential.

Keywords：pest detection；farmland environment；object detection；anchor-free；CenterNe

0 引言

农业作为国家经济的基石和社会稳定的关键支柱，其重要性不言而喻。（剩余15368字）