针对随机放置的马铃薯缺陷多项指标难以同时检测的问题，提出了一种基于高光谱信息融合的流形学习降维算法与极限学习机（extreme learning machine，ELM）相结合的方法，该方法可同时识别马铃薯的多项缺陷指标。分别采集发芽、绿皮、黑心和合格马铃薯的反射高光谱数据（390～1040 nm），在光谱维，提取马铃薯样本感兴趣区域（region of interest，ROI）的平均光谱，分别采用扩散映射（diffusion maps，DM）、局部线性嵌入（locally linear embedding，LLE）和海森局部线性嵌入（hessian locally linear embedding，HLLE）3种流形学习降维算法对光谱数据进行降维；在图像维，对马铃薯伪彩色图像进行形态学处理，获取基于灰度共生矩阵（gray level co-occurrence matrix，GLCM）的图像纹理信息，采用连续投影算法（successive projections algorithm，SPA）优选图像纹理特征；融合光谱维信息和图像维信息，分别建立基于极限学习机（ELM）与支持向量机（support vector machine，SVM）的马铃薯多分类识别模型。结果表明，扩散映射结合极限学习机（DM-ELM）模型的预测结果较优，该模型对发芽、绿皮、黑心和合格马铃薯样本的单一识别率分别为97.30%、93.55%、94.44%和100%，混合识别率达到96.58%，时间为0.11 s，可知高光谱信息融合技术结合流形学习降维算法可同时识别随机放置马铃薯的多种缺陷指标。%In order to conquer the difficulty of simultaneously recognizing the multiple defects of potatoes samples placed randomly, this paper proposed a non-destructive detection method which combined manifold learning dimension reduction algorithm based on hyperspectral information fusion and extreme learning machine (ELM) to simultaneously distinguish the multiple defects of potatoes. In this paper, 367 potatoes were picked which were made up of 111 sprouting potatoes, 90 green rind potatoes, 46 blackheart potatoes and 120 normal potatoes. The hyperspectral image acquisition system contained imaging spectroradiometer (SPECIM, V10E, Finland), data acquisition box, lighting system, electric moving stage and objective table. The reflection hyperspectral information of all those potatoes was acquired by using the hyperspectral image acquisition system, whose spectral wavelength ranged from 390 to 1040 nm. The hyperspectral information included the spectral information from 520 wave bands and the imaging information from 520 gray images. After the correction to hyperspectral data with the standard black and white board, hyperspectral data in the range of 450-990 nm, which had high signal-to-noise ratio (SNR), were selected as original spectrum for subsequent processing. To deal with the spectral information, the average spectrum was abstracted from the region of interests (ROI) on every potato sample by using the environment for visualizing images (ENVI). Comparing several data preprocessing methods, detrend was determined as the optimal spectral preprocessing method. Diffusion maps (DM), locally linear embedding (LLE) and hessian locally linear embedding (HLLE) were respectively utilized for the purpose of cutting down the dimension of spectrum data after the spectral preprocessing named detrend. To deal with the hyperspectral imaging information, every pseudo-color image of potatoes was morphologically processed before extracting 84 image texture characteristics based on gray level co-occurrence matrix (GLCM). By using successive projections algorithm (SPA), 10 texture features were properly selected, which consisted of homogeneity of 45° GLCM in R gray image, energy of 90° GLCM in R gray image, correlation of 90° GLCM in R gray image, contrast of 135° GLCM in R gray image, correlation of 0° GLCM in G gray image, contrast of 0° GLCM in G gray image, correlation of 90° GLCM in G gray image, homogeneity of 45° GLCM in B gray image, correlation of 90° GLCM in B gray image and contrast of 135° GLCM in B gray image. After the fusion of the spectral characteristics and image features, the 367 potato samples were divided into calibration set and test set. The calibration set had 250 potato samples including 74 sprouting potatoes, 59 green rind potatoes, 28 blackheart potatoes and 89 normal potatoes. The whole number of samples in test set was 117 which contained 37 sprouting potatoes, 31 green rind potatoes, 18 blackheart potatoes and 31 normal potatoes. Six models were respectively established based on support vector machine (SVM) and ELM, which were DM-SVM, LLE-SVM, HLLE-SVM, DM-ELM, LLE-ELM and HLLE-ELM. Comparing and analyzing the results of the 6 models and the spending time, we discovered that the DM was the best manifold learning dimension reduction algorithm to deduce the spectral information. We also found that the DM-ELM model was the best model because of the better recognition rate for recognizing the multiple defects of potatoes and spending less time. For the DM-ELM model, the single recognition rate of sprouting potatoes, green rind potatoes, blackheart potatoes and normal potatoes respectively reached 97.30%, 93.55%, 94.44% and 100%, and the mixed recognition rate reached 96.58%. It only cost 0.11 s to build the DM-ELM model. The results indicate that combining hyperspectral information fusion with manifold learning dimension reduction algorithm can simultaneously distinguish the multiple defects of potatoes.
展开▼
