摘要
Abstract
In order to achieve a non-destructive identification of textile fibers,a qualitative analysis model com-bining the Gram angle field and a lightweight convolutional neural network Mobile-V1,called GAF_Mobile-V1,is proposed.In order to solve the problem of absorption band overlap in infrared spectrometry,a Gram angle field is used to amplify the feature information of the original data and to enhance the feature representation of infrared spectral data.The Gram feature map is input to the Mobile-V1 model to realize the qualitative analysis,and comparison experiments are conducted with 1D-CNN and GAF_VGG19 models,combining the training curves and the accuracy of the test set to make a comparison:for the single-component qualitative analysis,the accuracy of the test set of GAF_Mobile-V1 is 94.6%,which is slightly lower than that of GAF_VGG19(95.4%)but much higher than the 40.2%of 1D-CNN;for the mixed-component qualitative analysis,the test set accuracy of GAF_Mobile-V1 is 92.3%,GAF_VGG19 is 91.4%,and 1D-CNN is 46.7%.And the feature visualization is used to intuitively react to the classification effect of the model.关键词
格拉姆角场/红外光谱/迁移学习/定性分析/轻量级网络Key words
Gram angle filed/infrared spectrum/transfer learning/qualitative analysis/lightweight network分类
轻工纺织
