支持向量机复合核函数的高光谱显微成像木材树种分类

采用体视显微高光谱成像方法，构建木材树种分类识别模型。利用SOC710VP体视显微高光谱图像采集系统获取可见光/近红外（372.53～1 038.57 nm）波段内的木材高光谱图像。首先，采用ENVI软件提取木材样本感兴趣区域（ROI）的平均光谱，分别采用连续投影算法（SPA）和竞争性自适应重加权算法(CARS)对光谱数据进行降维。再利用支持向量机（SVM）分别建立木材样本采集波段和特征波长下的分类模型。然后，在空间维采用第一主成分图像，计算基于灰度共生矩阵（GLCM）的木材纹理特征。在0°，45°，90°和135°四个方向计算能量、熵、惯性矩、相关性等16个特征参数后输入SVM进行木材树种分类处理。最后，采用四个复合核函数SVM进行光谱维和空间维的特征融合及分类识别。20个树种的分类实验结果表明，CARS的特征波长选择效果和运行速度较好一些，采用普通SVM进行木材光谱维特征分类处理时，测试集分类准确率达到了92.166 7%。采用基于GLCM的木材空间维纹理特征时，采用普通SVM的测试集分类准确率是60.333 0%，具有较低的分类精度。在将光谱维和空间维纹理特征进行数据融合及分类处理时，采用复合核函数SVM分类具有更好的效果。采用第二个复合核函数的SVM分类精度最高，测试集分类正确率是94.166 7%，运行时间为0.254 7 s。另外，采用第一个和第三个复合核函数的SVM的测试集分类准确率分别是93.333 3%和92.610 0%，运行时间分别为0.180 0和0.260 2 s。可以看出，采用这3种复合核函数的SVM进行木材树种分类，分类精度都高于采用普通SVM的光谱维或者空间维的分类识别精度。因此，利用体视显微高光谱成像和复合核函数SVM可以提高木材树种分类精度，为木材树种快速分类提供了参考。

Wood Species Recognition with Microscopic Hyper-Spectral Imaging and Composite Kernel SVM

In this paper, a stereomicroscopic hyper-spectral imaging scheme is used for wood species recognition. The SOC710VP hyper-spectral imaging system is used to pick up the wood images in visible and near-infrared spectral band (i. e., 372.53～1 038.57 nm). First, the ENVI software is used to pick up the mean spectra of wood sample’s Region of Interest (ROI). The Successive Projection Algorithm (SPA) and Competitive Adaptive Reweighted Sampling algorithm (CARS) are used for spectral dimension reduction. Second, a Support Vector Machine (SVM) is used to classify the wood samples in full spectral band and in feature wavelengths. Third, in the spatial dimension the 1st principal component image (PC1) is used to compute the wood texture features based on Gray Level Co-occurrence Matrix (GLCM). In the 4 directions of 0°, 45°, 90°, 135° the 16 feature parameters such as energy, entropy, inertia moment and so on are calculated and are put into SVM for wood species recognition. Lastly, the 4 composite kernels SVM are used to fuse the spatial-spectral features for wood species recognition. Experiments on 20 wood species classification indicate that CARS is a better choice in view of the feature wavelength selection and running speed and the classification accuracy for testing set reaches to 92.166 7% if the ordinary SVM is used for wood spectral classification. If the wood texture features based on GLCM are used, the classification accuracy for testing set reaches to 60.333 0% if the ordinary SVM is used. When the wood spectral and texture features are fused for classifications, the composite kernel SVM has the best classification accuracy. Especially, the classification accuracy of the 2nd composite kernel SVM is the highest with 94.1667% for testing set and a processing speed of 0.254 7 s. Moreover, the classification accuracy of the 1st or 3rd composite kernel SVM reaches to 93.333 3% or 92.610 0% with a running speed of 0.180 0 or 0.260 2 s. Therefore, wood species classification accuracy can be improved by use of hyper-spectral imaging and composite kernel SVM, which may be applied in the practical wood species classification system.