基于CNN一Transformer混合模型的辣椒病害识别

打开文本图片集

中图分类号：TP391.4；S436.418.1 文献标识码：A 文章编号：2095-5553（2025）10-0168-08

Abstract：To improve disease recognition accuracy and address the issues of classification errors and missed detections caused by inadequate feature extraction in traditional models，a CNN—Transformer hybridarchitecturewas proposed for pepper disease recognition，named Convolution-Transformer Fusion Network （CTF—Net）.In CTF—Net，a custom convolution module wasintroduced，named FeatureExtraction Convolution（FEC），at thelower layers to enhance low-levelfeatureextractionandreducethecomputational burdenon higher-level modules.Aditionally，the SEattntion mechanism is incorporated into the MV2block with paralel average polingand max pooling branches toadaptivelyadjust channel weights and enhance sensitivity tokey features.At higher layersof the network，adynamic feature fusion module （DCT）concatenates outputs from theconvolutionand Transformer modules，dynamically adjusting featurecontributions basedoninputdata characteristics.Thismechanismenablesthemodeltoeficientlyselectrelevantfeatures，eliminate redundancy，andoptimizefeatureintegration.Finally，asubsetofdiseaseimagesfromthePlant Villagedatasetis usedfor pre-training，followed by fine-tuning ona pepper disease-specific datasetto beter adapt the model tothe task，enhancing itsfeaturelearmingandgeneralizationcapabilities.Theexperimental resultsshow that theCTF—Netmodel，witha computational cost of only 640.6M ，achieved an identification accuracy of 97.5% after transfer learning on the pepper disease dataset.Compared with the classic models MobileViT，MobileNetV3—small，ResNet34，AlexNet，VGG16，and Swin Transformer， the classification accuracy was 7.6% ， 8.2% ， 6.6% ， 19.7% ， 3.7% ，and 5.3% higher， respectively.The model also performed better in precision，recall，specificity，and F1-score ：

Keywords：pepper; disease identification; convolutional neural network；self-atention mechanism;transfer learning

0 引言

随着全球农业现代化的不断推进，植物病害的识别与防治成为提高农作物产量和保障食品安全的重要课题[1]。（剩余13920字）