基于改进RT一DETR的棉田昆虫检测算法

打开文本图片集

中图分类号：S435.62；TP391.4 文献标识码：A 文章编号：2095-5553（2026）02-0210-08

Abstract：Given thepersistentchalengesof limitedaccuracy，frequentmisseddetections，andfalsepositivesincoton fieldinsect detection，this studyproposedanimprovedcotton field insectdetectionalgorithm based ontheRT—DETR framework.To adessthese issues，several key enhancements were introduced.First，WTConv wasused toreplace the secondconventionalconvolutionlayer intheresidual block.This modificationsignificantlyexpandedthemodel'sreceptive fieldwhile keepingthenumberof trainableparameterslow，therebyenhancing themodel’sabilitytodetectsmall targets efectively.Next，a dual-branch M2SA module was incorporatedto extract both global features and channel information. Thisimprovedthemodel'sunderstandingofcomplexfieldenvironmentsandboosteditsacuracyindetectingsmall insects.Aditionally，during the crossscale feature fusion stage，a Small Target Optimization Pyramid（STOP）was desiged toeficientlycapture and integratebothglobal and local features，further improvingthedetectionof small targets. Experimental results showed that the improved RT—DETR model achieved a mean Average Precision （mAP） of 95.4% ， which was8.9percentage points higher than thatof the original RT—DETR model.Furthermore，the enhanced model required only 12.1 million parameters and 42G of floating point operations per second，representing a 36% reduction in parameters and a 26% reduction in computational load compared to the original version. In summary，the improved RT— DETRmodel significantlyimproved theaccuracyandeficiencyof insect detectionincotton fields，offeringapractical precise solution for pest monitoring and management in agricultural applications.

Keywords：cotton field insects；target detection；RT—DETR；wavelet transform convolutic

0 引言

棉花虫害，尤其是棉铃虫、草蛉、中黑盲蝽等害虫，严重影响棉花的生长和产量。（剩余10695字）