基于目标检测的农业害虫计数方法综述

打开文本图片集

中图分类号：TP391.4 文献标识码：A 文章编号：2096-4706（2026）04-0060-08

Abstract： With the development of smart agriculture， Computer Vision methods powered by Deep Learning have been widelyapplied toagriculturalpest detectionandcounting.This paperreviews recentadvances in ObjectDetectionforthisfeld， and compares the diffrences between single-stage detectors （YOLOseries，SSD）and two-stage detectors （FasterR-CNN） in termsof precision，speed，anddeployment.Representativeimprovement strategiesandtheirperformanceondatasetssuchas AgriPest，IstgDatadst4reuedIddiioolyedeuatetrics（APrsieal， MAE， RMSE， R2 ） and their applicability to detection and counting tasks are summarized.The current challenges，including small object densityomplex backgrounds，data imbalane，andedgedeployment，arealsoconcluded.Finalyfuturedirectiossch as multimodal and spatiotemporal modeling，generative model-baseddata augmentation，unified evaluation frameworks，and lightweight optimizationareproposed，providingareference forinteligent agriculturalmonitoringresearchandapplications.

Keywords： agricultural pest; Object Detection; counting; Deep Learning; dataset; evaluation metrics

0 引言

农业害虫是威胁全球粮食安全和作物产量的重要因素，其高效监测与科学防控是农业可持续发展的关键环节[1-2]。（剩余14792字）