基于偏振成像技术的大米品种真实性识别  

Authenticity Recognition of Rice Varieties Based on Polarization Imaging Technology

作　　者：傅婕 吴跃 FU Jie;WU Yue(National Engineering Laboratory for Deep Processing of Rice and Byproduct,College of Food Science and Engineering,Central South University of Forestry and Technology,Changsha 410004,China)

机构地区：[1]中南林业科技大学食品科学与工程学院稻谷及副产物深加工国家工程实验室,湖南长沙410004

出　　处：《食品工业科技》2025年第7期235-247,共13页Science and Technology of Food Industry

摘　　要：为了开发一项无损、高效的大米品种真实性图像识别技术,本文以亲缘关系相近的3种粳米和3种籼米为研究对象,分别采集每种大米的可见光及其4种不同的偏振图像(即偏振强度I 0°图像、偏振Stokes矢量S0图像、偏振角图像、偏振度图像),结合6种卷积神经网络算法(即AlexNet、VGG16、GoogLeNet、ResNet34、DenseNet、ConvNeXt V2),分别建立基于不同图像类型和算法的粳米和籼米品种真实性识别模型。通过对比这些模型验证集的准确率发现,在粳米品种真实性识别中,基于ResNet网络的可见光图像,实现了最高的100%识别准确率;基于VGG16网络的偏振度图像,准确率达到98.5%。在籼米品种真实性识别中,基于VGG16网络的偏振Stokes矢量S0图像,最高实现了99.5%的准确率;基于VGG16和ResNet网络的偏振度图像均最高达到99.3%的准确率。本文证明了偏振成像技术应用在大米品种真实性识别中的现实可行性,也为粳米和籼米品种真实性识别时选择合适的图像类型和算法提供了重要参考依据。To develop a non-destructive and efficient image recognition technology to verify rice variety authenticity,this study centered on three closely related japonica rice varieties and three indica rice varieties.Visible light images and four distinct types of polarized images(polarization intensity I0°image,polarization Stokes vector S0 image,polarization angle image,and polarization degree images)were gathered for each rice type.Utilizing six convolutional neural network algorithms(AlexNet,VGG16,GoogLeNet,ResNet34,DenseNet,and ConvNeXt V2),models were established to identify the authenticity of japonica and indica rice varieties based on various image types and algorithms.When comparing the accuracy of these models on validation sets,it was observed that for identifying the authenticity of japonica rice varieties,the ResNet network based on visible light images achieved the highest accuracy at 100%,while the VGG16 network based on polarization degree images attained 98.5%.In the case of identifying the authenticity of indica rice varieties,the VGG16 network using polarization Stokes vector S0 images recorded the highest accuracy of 99.5%,while both the VGG16 and ResNet networks using polarization degree images achieved an accuracy of 99.3%.This study highlights the practical feasibility of employing polarization imaging technology for authenticating rice varieties and offers valuable reference data for selecting appropriate image types and algorithms for japonica and indica rice variety identification.

关 键 词：大米 品种 真实性识别 偏振成像技术 深度学习 

分 类 号：TS201.6[轻工技术与工程—食品科学]