基于脉冲驱动自注意力的害虫图像识别模型

打开文本图片集

中图分类号：TP391.41 文献标识码：A 文章编号：2095-5553（2025）12-0032-08

Abstract：Pestimagerecognitionisanimportantagricultural intelligentappication，whichcanhelpfarmers findand control pestsintimeandimprovecropyieldandquality.However，thetraditional imagerecognitionmodels basedon convolutionalneuralnetworksusuallrequirealotofcomputationalresourcesand energyconsumption，which limits their deploymentandaplicationonlow-powerdevices.Inorder tosolvethisproblem，apest imagerecognitionmodel based on pulse-driven self-atention （SDSA）is proposed，which uses the characteristics of pulsed neural network（SNN）to transform images into pulse sequences，and then extractsand classifies them byself-atention mechanism．Themodel in this paper has the follwing four unique characteristics：（1）Event driven，no calculation is triggeredwhen the inputis zero.（2）Binary pulsecommunication，allmatrix multiplications related tothe pulse matrix can beconverted to sparse addition：（3）Self-atention with linearcomplexityin both taskandchanneldimensions.（4）Inpulse-form queries，the operationbetween key and valuevectors is mask andaddition，andonly includes sparse adition operation，which greatly reducesthecomputational loadand energyconsumption.Experiments were carred outonpest image datasets，and the results show that the proposed model accuracy rate can reach 93.56% ，which is beter than the existing deep learning model，and at the same time，it has lower computational cost and energy consumption.

Keywords：pulse drive；self-attention model;pest image recognition；smart agriculture；light weight

0 引言

随着全球人口增长和食品需求增加，农业生产持续面临着提高产量和质量的压力[1-3]。（剩余13506字）