Rapid Qualitative Analysis of Wool Content Based on Improved
U-Net++ and Near-Infrared Spectroscopy
LENG Si-yu1, 2, QIAO Jia-hui1, WANG Lian-qing3, WANG Jun1, 2*, ZOU Liang1
1. Engineering Research Center of Intelligent Control for Underground Space, Ministry of Education, China University of Mining and Technology, Xuzhou 221116, China
2. School of Information and Control Engineering, China University of Mining and Technology, Xuzhou 221116, China
3. Xuzhou Zhaoheng Industrial Control Technology Co., Ltd., Xuzhou 221008, China
摘要： 羊毛制品因其柔软、保暖性好等优点广受欢迎，羊毛含量是衡量这类产品质量的重要依据。目前市场上羊毛制品质量参差不齐，传统检测方法具有破坏性大、主观性强等缺点，已无法满足实时快速评估目标羊毛制品质量情况的需求。近红外光谱技术是一种无需破坏样品结构、可模型封装操作的快速测量方法。将近红外光谱技术和深度学习技术融合，提出了一种基于注意力机制和U-Net++网络的羊毛含量快速定性分析方法。在数据准备方面，使用手持便携式光谱仪采集羊毛制品样本的光谱数据，其波段范围为908.1~1 676.2 nm，并根据其含量的不同对原始样本进行了等级划分。为减少光谱采集方式对建模数据集的影响，针对同一样本在距探头5，6，8，9和19 mm 5种高度，分别采集了5次光谱数据，并使用马氏距离法剔除异常样本，最终共5 125组光谱数据用于建模。在模型选择方面，U-Net++网络可通过下采样、跳跃连接和上采样等环节实现对光谱数据的特征提取，并进一步对样本进行分类预测。然而，该网络使用了大量密集的跳跃连接，易产生模型参数冗余、低层特征被重复使用等问题。鉴于此，在原始网络的基础上引入了注意力门控模块，可以更有效地提取特征信息，提高预测精度。建模过程中，将羊毛制品样本90%的光谱数据用于训练和验证，10%的数据用于测试。实验结果表明，基于U-Net++网络的预测模型在独立测试集上的准确率为93.59%，召回率为93.53%，精确率为94.24%，其性能超过了多种传统分类模型。同时，与U-Net，Attention U-Net等其他U-Net系列网络模型相比，所提出的分类模型的各项评价指标也高于上述网络模型，验证了跳跃连接和注意力门控模块的效果。基于近红外光谱技术，构建了Attention U-Net++模型，为羊毛含量快速无损检测提供了一种新思路，具有一定的理论和实际意义。
关键词：近红外光谱；深度学习；羊毛制品；U-Net++
Abstract：Wool products are popular because of their softness and warmth. The content of wool is an important indicator of the quality of wool products. However, the quality of wool products in the market varies. In addition, traditional testing methods are destructive, and the results might be subjective, which can no longer meet the need to evaluate the quality of the target wool products quickly. NIR spectroscopy is a rapid measurement method that does not require the destruction of sample structure and can be embedded with machine learning models. Because of this, this paper proposes a rapid qualitative wool content evaluation method via fusing NIR spectroscopy and attention-based U-Net++. In terms of data preparation, this paper employs a handheld portable spectrometer to collect spectral data of wool product samples with a wavelength range of 908.1 to 1 676.2 nm. The original samples are graded according to their contents. The experiments collected spectral datasets of the same sample at 5 heights of 5, 6, 8, 9 and 19 mm from the spectrometer, and abnormal samples were removed by Mahalanobis distance. 5 125 sets of spectral data were used for the final data modeling. Regarding model selection, the U-Net++ network provides an end-to-end way for feature extraction and classification with down-sampling, jump connections and up-sampling operations. However, due to alarge number of skip connections, it reuses low-level features, and the models might contain redundant parameters. This paper introduces an attention-gating module which can extract feature information more effectively and improve prediction accuracy. The spectral data corresponding to 90% of wool product samples is used for training and validation, and the rest spectral data is used for testing. The experimental results show that the prediction model based on the U-Net++ network obtains an accuracy of 93.59%, a recall of 93.53%, and a precision of 94.24% on the independent test set, all of which outperform traditional classification models. Meanwhile, the classification model proposed in this paper outperforms other U-Net series networks, such as U-Net and Attention U-Net, demonstrating the effectiveness of the skip connection and attention-gating modules. In this paper, the spectral analysis based on the Attention U-Net++ model and portable near-infrared spectrometer provides a practical and meaningful way for rapid, nondestructive inspection of wool content.
Key words：Near-infrared spectrum; Deep learning; Wool products; U-Net++
冷思雨，乔嘉慧，王连庆，王 军，邹 亮. 基于改进U-Net++和近红外光谱技术的羊毛含量快速定性分析方法研究[J]. 光谱学与光谱分析, 2023, 43(01): 303-309.
LENG Si-yu, QIAO Jia-hui, WANG Lian-qing, WANG Jun, ZOU Liang. Rapid Qualitative Analysis of Wool Content Based on Improved
U-Net++ and Near-Infrared Spectroscopy. SPECTROSCOPY AND SPECTRAL ANALYSIS, 2023, 43(01): 303-309.
[1] Xing W Y，Xin B J，Deng N，et al. Measurement，2019，132：11.
[2] Celio P. Analytica Chimica Acta，2018, 1026: 8.
[3] Li J，Zhang H，Zhan B，et al. Infrared Physics & Technology，2019，104：103154.
[4] Chen H，Chen A，Xu L，et al. Agricultural Water Management，2020，240：106303.
[5] Be K，Grabska J，Huck C. Journal of Pharmaceutical and Biomedical Analysis，2021：193: 113686.
[6] Weng S，Guo B，Tang P，et al. Spectrochimica Acta Part A: Molecular and Biomolecular Spectroscopy, 2020，230：118005.
[7] Zou L，Yu X，Li M，et al. IEEE Transactions on Instrumentation and Measurement，2020，69（10）：8043.
[8] Sun X D，Zhu K. International Journal of Clothing Science and Technology，2019，31（3）：326.
[9] Liu Y S，Zhou S B，Liu W X，et al. Journal of Near Infrared Spectroscopy，2018，26（1）：34.
[10] Chen H，Tan C，Zan L. Spectrochimica Acta Part A: Molecular and Biomolecular Spectroscopy，2018，201：229.
[11] Zhou J F，Wang R W，Wu X Y，et al. Applied Spectroscopy，2017，71（10）：2367.
[12] Sun X T，Yuan H F，Song C F，et al. Microchemical Journal，2019，149：104018.
[13] Fang X，Gao T，Zou L，et al. Sensors，2020，20（23）：6784.
[14] Jiao L C，Zhao J. IEEE Access，2019，7：172231.
[15] Lei M，Li J，Li M，et al. Diagnostics，2021，11（3）：534.
[16] Chen H Z，Chen A，Xu L L，et al. Agricultural Water Management，2020，240：106303.
[17] Yu G W，Ma B X，Chen J C , et al. Journal of Food Process Engineering，2021, 44（1）：e13602.
[18] Zhou Z，Siddiquee M，Tajbakhsh N，et al. IEEE Transactions on Medical Imaging，2020，39（6）：1856.
[19] Weng Y，Zhou T B，Li Y J，et al. IEEE Access，2019，7：44247.
[20] Lei M，Rao Z，Wang H，et al. Fuel，2021，294（2）：120475.
