基于双路多尺度卷积的近红外光谱羊绒羊毛纤维预测模型  

Near-Infrared Spectral Prediction Model for Cashmere and Wool Based on Two-Way Multiscale Convolution

作　　者：陈锦妮[1] 田谷丰 李云红[1] 朱耀麟[1] 陈鑫 门玉乐 魏小双 CHEN Jin-ni;TIAN Gu-feng;LI Yun-hong;ZHU Yao-lin;CHEN Xin;MEN Yu-le;WEI Xiao-shuang(School of Electronic Information,Xi'an Engineering University,Xi'an 710600,China)

机构地区：[1]西安工程大学电子信息学院,陕西西安710600

出　　处：《光谱学与光谱分析》2025年第3期678-684,共7页Spectroscopy and Spectral Analysis

基　　金：国家自然科学基金项目(62203344);陕西省科技计划项目(2022GY-053);陕西省自然科学基础研究重点项目(2022JZ-35);陕西省教育厅产业用纺织品协同创新中心重点研究项目(20JY026);陕西省教育厅产业化培育项目(23JC031);西安市科技计划项目(23DCYJSGG0008-2023);榆林市科技计划项目(CXY-2020-052)资助。

摘　　要：羊绒具有轻盈舒适、光滑柔软、稀释透气以及保暖好的特点,由于羊绒价格十分昂贵,因此市场上的羊绒产品质量良莠不齐。现有的显微镜法、DNA法、化学溶解法和基于图像的方法具有损坏样本、设备昂贵、主观性强等不足。近红外光谱技术是一种非破坏性、可进行建模操作的快速测量方法。针对传统的建模方法通常无法学习出通用的近红外光谱波段特征,导致泛化能力弱,且羊绒羊毛纤维的近红外光谱波段特征相似,难以区分的问题,本文提出一种基于双路多尺度卷积的近红外光谱羊绒羊毛纤维预测模型。采集了羊绒羊毛样品的近红外光谱波段数据共1170个进行验证,近红外光谱波段数据范围是1300~2500 nm。利用两个并行卷积神经网络来提取近红外光谱波段的特征,采用原始近红外光谱波段数据和降维近红外光谱波段数据同时输入的方式,并利用多尺度特征提取模块进一步提取中间具有贡献力的近红外光谱波段特征,利用路径交流模块用于两路近红外光谱波段特征的信息交流,最后利用类级别融合得到羊绒羊毛纤维预测结果。在实验过程中,将采集的80%近红外光谱波段数据用于模型训练,20%近红外光谱波段数据用于模型测试。模型测试集的平均预测准确率为94.45%,与传统算法中的随机森林、SVM、1D-CNN等算法相比较分别提升了7.33%、5.22%、2.96%,并进行消融实验对所提模型的结构进一步验证。实验结果表明,本文提出的双路多尺度卷积的近红外光谱羊绒羊毛纤维预测模型可实现羊绒羊毛纤维的快速无损预测,为近红外光谱羊绒羊毛纤维预测提供了新的思路。Cashmere is characterized by lightness and comfort,smoothness and softness,dilution and breathability,and good warmth.Because it is very expensive,the quality of cashmere products in the market is mixed.Existing microscopy,DNA,chemical dissolution,and image-based methods have shortcomings such as damaged samples,expensive equipment,and high subjectivity.NIR spectroscopy is a rapid measurement method that is non-destructive and allows for modeling operations.Aiming at the problems that traditional modeling methods usually fail to learn universal near-infrared spectral band features,resulting in weak generalization ability,and that the near-infrared spectral band features of cashmere wool fibers are similar and difficult to distinguish,this paper proposes a near-infrared spectroscopy cashmere wool fiber prediction model based on two-way multi-scale convolution.In terms of data preparation,a total of 1170 near-infrared spectral band data of the original cashmere wool samples are collected for validation,and the range of the near-infrared spectral band data is 1300~2500 nm;in terms of model design,two parallel convolutional neural networks are utilized to extract the features of the near-infrared spectral band,and both the original near-infrared spectral band data and the downscaled near-infrared spectral band data are used as simultaneous.The original near-infrared spectral band data and the downscaled near-infrared spectral band data are input simultaneously.The intermediate contributing near-infrared spectral band features are further extracted using the multi-scale feature extraction module,and the path exchange module is used for the information exchange of the two near-infrared spectral band features.Finally,the cashmere wool fiber prediction results are obtained using the class-level fusion.In the experimental process,80%of the collected near-infrared spectral band data are used for model training and 20%of the near-infrared spectral band data are used for model testing.The average prediction accuracy of the t

关 键 词：羊绒羊毛 近红外光谱 深度学习 双路多尺度卷积神经网络 

分 类 号：O657.3[理学—分析化学]