光谱分析在木材种类识别领域具有一定的潜力, 而其中的中红外光谱也广泛应用于定性及定量分析。 该研究中红外光谱法识别木材种类的报道中, 基于深度卷积神经网络可以在参数较少的条件下获得较高的识别率。 提出了一种聚类分析(CA)、 对称点阵图像分析(SDP)与深度学习(DenseNet)相融合的算法, 利用DenseNet的优势提高中红外光谱法识别木材的准确率。 首先, 采集了愈疮木、 巴里黄檀、 刺猬紫檀、 大果紫檀、 螺穗木5种木材样本的250组中红外光谱数据, 应用欧式距离进行了异常值剔除, 剩余240组作为待分析数据并对其进行分类的可行性分析。 对光谱数据进行SDP转化分析, 确定SDP转化的最优参数; 之后, 运用CA筛选原始光谱数据的特征, 根据CA不同阙值确定了三组维数的特征并进行讨论, 通过对比三组特征数据经过SDP转化后图像间的类内相似性和类间差异性, 初步确定了最优维数特征; 再将初步确定的最优维数特征数据输入到SDP-DenseNet模型中, 获得模型识别的准确率; 最后通过对比分析验证了模型的有效性, 一方面将原始数据及其余两组对照维数的特征数据分别输入到SDP-DenseNet模型中, 对比识别的准确率; 另一方面以最优维数特征数据输入到随机森林中进行识别, 对比传统机器识别与SDP-DenseNet算法识别的准确率。 结果表明: 经CA特征筛选的SDP-DenseNet模型普遍高于原始数据直接输入到SDP-DenseNet模型的准确率, CA特征筛选最优维数为255维, 其测试集最高识别率达到了88.67%, 而对照组107维为77.78%, 322维为68.89%, 原始数据的SDP-DenseNet模型识别率仅为57.78%; 经CA特征筛选的最优维数数据对应的随机森林模型识别率较低, 仅为66.67%。 因此, 提出的CA-SDP-DenseNet模型能有效提高中红外光谱法识别木材种类的精度。

Spectral analysis technology has a certain potential in wood species identification, and mid-infrared spectroscopy technology is also widely used in qualitative and quantitative analysis. This research focuses on the identification of wood species by mid-infrared spectroscopy. Based on a deep convolutional neural network, an algorithm that combines cluster analysis (CA), symmetrical lattice image analysis (SDP) and deep learning (DenseNet) is proposed to achieve a high recognition rate with few parameters. With the advantages of DenseNet, the accuracy of wood recognition in mid-infrared spectroscopy is improved. First, 250 sets of mid-infrared spectroscopy data, including guaiacum sanctum, dalbergiabariensis, pterocarpuserinaceus, pterocarpusmacarocarpus, and spiraea, are collected. Through eliminating outliers based on Euclidean distance, the feasibility analysis of the remaining 240 groups as data to be analyzed and classified. The optimal parameters of SDP conversion are determined through the SDP conversion analysis of the original spectral data. The characteristics of original spectral data are filtered out through CA. According to CA, different thresholds determine the characteristics of the three groups of dimensions and related discussions are carried out. The optimal dimensional feature is initially determined by comparing the three sets of feature data, including the intra-class similarity and the inter-class difference between the images after SDP conversion. The determined optimal dimensional feature data is input into the SDP-DenseNet model to obtain model recognition accuracy. Finally, the comparative analysis verifies the validity of the model. On the one hand, the original data and the feature data of the other two sets of contrast dimensions are input into the SDP-DenseNet model to compare recognition accuracy. On the other hand, the optimal dimensional feature data is input into the random forest for recognition to compare the accuracy of traditional machine recognition and SDP-DenseNet algorithm recognition. According to the results, the accuracy of the SDP-DenseNet model filtered by the CA feature is generally higher than that of the SDP-DenseNet model directly input to the original data. The optimal dimension of CA feature selection is 255 dimensions, with the highest recognition rate of 88.67%. In the control group, the recognition rate of 107 dimensions is 77.78%, and the recognition rate of 322 dimensions is 68.89%. In contrast, the SDP-DenseNet model recognition rate of the original data is only 57.78%. The recognition rate of the random forest model corresponding to the optimal dimensionality data screened by clustering features is relatively low, only 66.67%. Therefore, the CA-SDP-DenseNet model proposed in this study can effectively improve the accuracy of mid-infrared spectroscopy in identifying wood species.