•  
  •  
  •  
  •  
  •  
  •  
  •  
  •  
  •  
期刊检索


检索
检索项:
检索词:
总目录
  • WCSB9︱2019(第九届)世界采样和混样大会 查看全文>>
  • 关于召开第十五届全国青年分析测试学术报告会的通知 查看全文>>
友情链接
“动态”近红外光谱结合深度学习图像识别和迁移学习的模式识别方法研究
    点此下载全文
作者单位
孙禧亭,袁洪福,宋春风 北京化工大学材料科学与工程学院 
基金项目:国家重点研发计划(2016YFF0102504)
中文摘要:该文以山羊绒与山羊绒/羊毛混纺织物以及纯棉与丝光棉织物为研究对象,使用其“动态”光谱,扩大类间的光谱差异信息,通过融合其同步和异步二维相关光谱,用多张动态光谱构造一张能反映细节化学差异信息的“化学图像”。使用GoogLeNet深度神经网络图像识别模型结合迁移学习,建立了一种光谱分类的新方法。收集了234个织物样品,制备水含量分别为0、5.4%、11.2%和16.3%的样本,同时采集样品的漫反射近红外光谱。使用干基样品的多种预处理光谱,利用线性分类方法簇类独立软模式识别(SIMCA)和非线性方法支持向量机(SVM),共建立了16个分类模型。其中,山羊绒与山羊绒/羊毛混纺织物的SIMCA和SVM最优预测正确率分别为63.33%和70.09%,纯棉与丝光棉织物的分别为71.02%和72.51%,均不能实现有效分类。新方法对山羊绒与山羊绒/羊毛混纺织物的预测正确率为92.59%,纯棉与丝光棉织物的为94.74%,获得了有效分类。该文首次将图像分类方法用于光谱分类识别,开辟了一种新的研究途径。针对实际应用能收集到的样品属于小样本,不能满足深度学习需要大数据样本的问题,使用迁移学习方法使深度学习框架适应了光谱分类(小样本),为人工智能领域中先进的识别技术用于解决化学问题提供了一个成功示范。
中文关键词:近红外光谱  模式识别  深度学习  迁移学习
 
Study on Pattern Recognition Method Using ‘Dynamic’ NIR Spectroscopy with Deep Learning based Image Identification and Transfer Learning
Abstract:A new classification method based on chemical image was established using ‘dynamic’ near infrared(NIR) spectroscopy with a deep learning based image recognition model GoogLeNet and transfer learning,with cashmere,cashmere/wool blends textiles,cotton and silk cotton textiles as the targets.Moisture perturbation was proposed to apply in this paper,collecting ‘dynamic’ spectra,expanding the spectral differences between samples of different types,and thus fusing the synchronous and asynchronous two dimensional map of dynamic spectra into a ‘chemical image’ which reflects the detailed differences between samples.A total of 234 textile samples were collected,and the samples with water contents of 0,5.4%,11.2% and 16.3% were prepared.Several preprocessing methods were employed before modeling.A total of 16 classification models were established,in which the best SIMCA and SVM models for cashmere vs cashmere/wool blends have the accuracies of 63.33% and 70.09%,while those of cotton and silk cotton textiles are 71.02% and 72.51%,respectively.Results demonstrated that the developed method is effective,the overall prediction correct rates of models are 92.59% for cashmere and blended and 94.74% for cotton and silk cotton.This contribution provides a successful demonstration for advanced identification techniques in the field of deep learning for solving chemical problems.
Key Words:near infrared spectroscopy  pattern identification  deep learning  transfer learning
引用本文:孙禧亭,袁洪福,宋春风.“动态”近红外光谱结合深度学习图像识别和迁移学习的模式识别方法研究[J].分析测试学报,2020,39(10):1247-1253.
摘要点击次数: 2191
全文下载次数: 511
查看全文  下载PDF阅读器