基于多模态融合的玉米种子成熟度的无损检测

Non-destructive detection of Zea mays L. seed maturity based on multimodal fusion

中文关键词:  高光谱成像  玉米种子成熟度  多模态融合  特征波长提取  纹理特征提取

英文关键词:hyperspectral imaging  Zea mays L. seed maturity  multimodal fusion  characteristic wavelength extraction  texture feature extraction

中文摘要:

      目的 应用高光谱成像技术, 结合多模态融合方法, 实现对玉米种子成熟度精准、无损检测。方法 获取高、低成熟度玉米种子高光谱图像, 采用自举软收缩算法与连续投影算法的级联算法(bootstrapping soft shrinkage-successive projections algorithm, BOSS-SPA)进行特征波长提取, 采用灰度共生矩阵法(gray-level co-occurrence matrix, GLCM)进行图像纹理特征提取, 选择能量、熵、相关性、逆方差和对比度5个特征参数, 将光谱与图像数据进行特征级融合, 利用偏最小二乘判别(partial least squares-discriminant analysis, PLS-DA)和最小二乘支持向量机(least squares support vector machine, LS-SVM)建立玉米种子成熟度分类模型。结果 确定使用SG卷积平滑-标准正态变量变换(Savitzky-Golay convolution smoothing-standard normal variable, SG-SNV)作为最佳光谱预处理方法, 采用BOSS-SPA方法提取的11个波长表现出良好建模性能, 基于光谱图像融合数据的模型测试集总体识别准确率均达到95%以上。结论 高光谱技术结合多模态特征融合方法有望成为玉米种子成熟度的无损检测提供切实可行的参考方法。

英文摘要:

      Objective To achieve accurate and non-destructive detection of Zea mays L. seed maturity by applying hyperspectral imaging technology combined with multimodal fusion methods. Methods Hyperspectral images of high and low maturity Zea mays L. seeds were acquired. The cascade algorithm of bootstrapping soft shrinkage and successive projections algorithm (BOSS-SPA) was used for feature wavelength extraction, while the gray-level co-occurrence matrix method (GLCM) was used for image texture feature extraction. Five feature parameters—energy, entropy, correlation, homogeneity and contrast were selected to integrate the spectra with the image data in a feature level fusion. Results The partial least squares-discriminant analysis (PLS-DA) and least squares support vector machine (LS-SVM) were used to establish a Zea mays L. seed maturity classification model. The use of Savitzky-Golay convolution smoothing-standard normal variable transformation (SG-SNV) was identified as the best spectral preprocessing method, and the 11 wavelengths extracted using the BOSS-SPA method showed good modelling performance, and the overall recognition accuracies of the model test set based on the fused data of the spectral images all reached over 95%. Conclusion Hyperspectral technology combined with multimodal feature fusion method is expected to provide a practical reference method for non-destructive detection of Zea mays L. seed maturity.