基于改进MobileNetV2的烟丝种类识别

打开文本图片集

中图分类号：TS493 文献标识码：A 文章编号：2095-5553（2025）08-0058-08

Abstract：To address theissue of smallandindistinguishable tobacco strands morphology among different types，a tobacco strandsclasification method basedon the improved MobileNetV2 was proposed in this study.The method employedMobileNetV2as the base network and introduceda multi-scale feature fusion moduletocapturerichdetailsof tobaccostrands.Excessive bottlenecks inthe backbone network wereremoved，and the clasifierwasredesigned toreduce networkdepth.Knowledgedistilationtechniqueswereincorporated，utilizingatransfer-learnedResNet5Onetwork to guide the trainingof themodified MobileNetV2 for lightweight model implementation.Experimentalresults demonstrated that the improvedMobileNetV2-based tobacco strands clasification method achieved an accuracyof 95.37% in recognizing various tobacco strands types，showing an 8.6% improvement over the baseline network. The parameter count was reduced to 0.62M ，a decrease of 1.61M compared to the baseline network.Furthermore，when compared to traditional classificationnetworks（GoogLeNet，AlexNet，ResNet50，VGG16），the proposed method exhibited higher accuracy in tobacco strands recognition with lower computational complexity.

Keywords：tobacco strands identification；deep learning；convolutionalneural network；knowledgedistilation；lightweight

0 引言

随着各类烟草制品的需求量逐年攀升，烟草已经成为全球性产业。（剩余12551字）