基于双分支特征提取的害虫分类方法研究

打开文本图片集

中图分类号：S43；TP391.4 文献标识码：A 文章编号：2095-5553（2026）03-0222-07

Research on insect classification method based on dual-branch feature extraction

Chen Yuefeng，Gao Xindan （College of Com puter and Control Engineering，Northeast Forestry University，Harbin，150006，China）

Abstract：To overcome the difficulties of accuratepest classificationcausedbycomplex bodystructures，diverse poses，and intricate backgroundsof pests，this studyproposedanewpestclasification model basedonadual-branch featureextractionand fusion framework.First，to effectively capture local features of pests with varying postures andcomplex structural details，aconvolutional neural network（CNN）branch incorporating aspatial transformation module was designed. Second，a Transformer branch was employed to extract global contextual features from pest images with complexbackgrounds..Finally，a featurefusion module integrates thecomplementarylocal and global features from both branches to enablerobust pest classification.The proposed model was evaluated on three datasetswith different characteristics， achieving classification accuraciesof 74.01% ， 98.21% ，and 90.12% allofwhich outperformed mainstream baseline methods.The resultsdemonstrated theefectiveness of incorporating the CNN branch and the feature fusion mechanism，which enhancedthe Transformer'sability to learnmore comprehensivefeatures， thereby addressing the challenge of limited classification accuracy in pest identification.

Keywords：pestclassification；self-attention；convolutional neural network；spatial transformer network；feature fusion

0 引言

在自然生态系统中，害虫的侵害严重影响了农林作物的正常生长1，准确、及时地对害虫进行识别，能够降低损失。（剩余11723字）