Nondestructive Identification of Millet Varieties Using Hyperspectral Imaging Technology
Nondestructive Identification of Millet Varieties Using Hyperspectral Imaging Technology
复制标题
DOI：
10.1007/s10812-020-00962-y
复制
发表时间：
2020-03-26
影响因子：
0.7
通讯作者：
Wang, W.
中科院分区：
文献类型：
--
作者：
Wang, X.;Li, Z.;Wang, W.
在这项研究中，我们使用了 8 个小米品种，并对 480 个小米样本拍摄了可见光-近红外高光谱图像。从高光谱图像中提取小米样品的光谱和图像特征，包括纹理和颜色特征。利用提取的光谱和图像特征建立了小米品种识别的支持向量机（SVM）模型。引入具有注意力机制的注意力-卷积循环神经网络（attention-CRNN）模型用于谷子品种识别，利用图像与光谱特征融合的方法建立了谷子品种识别的SVM和attention-CRNN模型。我们发现最高的数学变换方法是倒对数方法。倒数对数光谱特征曲线SVM品种分类模型的识别准确率为73.13%。利用图像特征的SVM模型对8个小米品种的整体识别准确率仅为61.25%。采用图像与光谱信息融合方法的SVM模型识别精度大幅提高，整体准确率达到77.5%，谷子品种的最低判别准确率从50%提高到65%。 Attention-CRNN模型的总体识别准确率为87.50%，比SVM模型提高了10%，谷子品种的最小判别准确率从65%提高到90%。结果表明，attention-CRNN模型提高了8个小米品种的整体识别准确率，并且大幅度提高了最小识别准确率。注意力 CRNN 模型在小米和其他小谷物品种的无损识别方面显示出巨大的潜力。
In this study, we used eight millet varieties and took visible-near-infrared hyperspectral images of 480 millet samples. Spectral and image characteristics, including texture and color features, of the millet samples were extracted from the hyperspectral images. Support vector machine (SVM) models for millet variety identification were established using the extracted spectral and image characteristics. An attention-convolutional recurrent neural network (attention- CRNN) model with attention mechanism was introduced for the identification of millet varieties, and the SVM and attention- CRNN models for millet variety identification were established using an image and spectral features fusion method. We found that the highest mathematical transformation method was the reciprocal logarithmic method. The identification accuracy of the SVM cultivar classification model with the reciprocal logarithmic spectral characteristics curve was 73.13%. The overall identification accuracy of the SVM model for the eight millet varieties using the image features was only 61.25%. The identification accuracy of the SVM model using the image and spectral information fusion method greatly improved the overall accuracy rate to 77.5%, and the minimum discrimination accuracy of the millet varieties increased from 50 to 65%. The overall identification accuracy of the attention-CRNN model was 87.50%, which is 10% higher than that of the SVM model, and the minimum discrimination accuracy of the millet varieties increased from 65 to 90%. The results show that the attention-CRNN model improved the overall identification accuracy of the eight millet varieties and greatly improved the minimum identification accuracy. The attention-CRNN model shows great potential for the nondestructive identification of millet and possibly other small grain varieties.