基于频域SwinTransformer的植物叶片病害识别研究

打开文本图片集

关键词：植物叶片;病害识别;数据增强;频域SwinTransformer;边缘检测;高斯滤波中图分类号：TP391.41；S4 文献标识码：A 文章编号：2095-5553（2025）10-0128-10

Abstract：Plantdiseasesand pests poseserious threats toagricultural production，andneedtobemonitoredand prevented inatimelymanner.Dueto the vast varietyof plant diseasesand pests and their similar earlysymptoms，it is extremely chalenging foragricultural workers todiferentiate betweenthem.Forthisreason，this studyproposesa plantleaf disease identificationmethodbasedonthefrequency-domain Swin Transformer.Initiall，themodel’strainingeficiency was enhancedbyusing animproved CutMix dataaugmentationalgorithm，which focuses themodelon thecritical partsof diseaseimages，alowing ittolearn more information，avoid overfiting，and improvegeneralization performance. Subsequently，Gausian filtering and edge detection were employed to reduce the negative impact of background noiseon theaccuracyofdiseaserecognition，highlighting theleafcontour information.Finally，afrequency-domainlayer wasadded tocapture local featuresof disease images.Theexperiments demonstrate that theproposedmethodhasachieved accuracies of 98.59% ， 100% and 99.58% on datasets for tomatoes，riceand cotton，respectively，which represent improvements of 1.34% ， 0.12% and 0.5% over the previous methods.Additionally，the detection speeds are increased by 2.54 frames/s，4. O4 frames/s and 9.97 frames/s，respectively.

Keywords：plant leaf；disease identification；dataaugmentation；frequency domain Swin Transformer;edge detection; gaussian filtering

0 引言

随着全球竞争加剧和气候变化，农作物的安全问题显得尤为重要。（剩余15502字）