基于介电频谱技术的苹果品种识别  

Apple variety identification by dielectric spectrum technology

作　　者：郝玉梅 花元涛 李文凤 刘清和 HAO Yumei;HUA Yuantao;LI Wenfeng;LIU Qinghe(College of Information Engineering,Tarim University,Alar,Xinjiang 843300;Key Laboratory of Tarim Oasis Agriculture Tarim University,Ministry of Education.Alar,Xinjiang 843300)

机构地区：[1]塔里木大学信息工程学院,新疆阿拉尔843300 [2]塔里木绿洲农业教育部重点实验室塔里木大学,新疆阿拉尔843300

出　　处：《塔里木大学学报》2024年第3期98-106,共9页Journal of Tarim University

基　　金：塔里木大学校长基金硕士项目(TDZKSS202131);塔里木大学大学生创新创业训练项目(22000033635)。

摘　　要：为快速、精准、无损地识别苹果品种,促进苹果产业快速发展,凸显高品质品种苹果优势,本研究采用介电频谱技术,以阿克苏地区3个品种的‘红富士’苹果作为研究对象,基于LCR数字测试仪采集了120个苹果样本在0~100 kHz的介电特频谱数据作为原始输入参数,在全频和连续投影算法优选频率条件下,利用误差反向传播网络与极限学习机2种方法建立了品种识别模型,并对模型精度进行了分析比较。结果表明,所建立的模型平均准确率均在80%以上,频率优选条件下的2种模型分类准确率可达到90%。然而由于以上模型方法存在无用信息,导致建立的模型稳定性较差,需要数据预处理、降维等步骤,操作耗时,且过程繁琐,致使分类结果稳定性较差。基于此,设计了一维卷积神经网络品种分类模型,与其他模型相比,一维卷积神经网络分类模型以原始参数作为输入,校正集与预测集中的平均分类准确率分别为98.48%和99.26%,模型稳定性更好,且可简化模型复杂度,改善苹果分类的准确性和稳定性,更适宜于苹果品种的识别。In order to quickly,accurately and non-destructive identify apple varieties,promote the rapid development of the apple industry,and highlight the advantages of high-quality apple varieties,this paper adopted dielectric spectrum technology for apple variety identification.Three varieties of‘Red Fuji’apples from Aksu region were selected as the research objects,and 120 apple samples were collected using LCR digital testing equipment with dielectric spectrum data at 0-100 kHz as the original input parameters.Under the optimal frequency conditions of full frequency and continuous projection algorithm,a variety recognition model was established by using two methods:error backpropagation network and extreme learning machine.Further,the accuracy of the methods was analyzed and compared.The results show that the average accuracy of the established models is above 80%,and the classification accuracy of the two models under frequency optimization conditions can reach 90%.However,the above model methods contain redundant information which undermines the stability of the model.Based on this,an one-dimensional convolutional neural network variety classification model was designed.Compared with other models,the new model uses raw parameters as input,and the average classification accuracy in the correction set and prediction set is 98.48%and 99.26%,respectively.The new model has better stability,simplifies model complexity,improves the accuracy and stability of apple classification,and is more suitable for identifying apple varieties.

关 键 词：品种识别 阿克苏苹果 介电频谱 特征频率 分类模型 

分 类 号：S23-0[农业科学—农业机械化工程]