基于光谱和图像特征的阔叶木材与针叶木材同时分类算法研究

木材是人们生活中必不可少的可再生资源，同时在建筑、工艺、家具、结构材料等方面有着举足轻重的地位。市场中常见的木材品种繁多，其品质和价格千差万别，使用智能化技术对木材进行正确的分类不仅可以防止不法商贩“以次充好”，也可以大幅度降低木材分类人员的工作难度。通过木材的遗传信息和解剖学信息可以得到较为准确的木材分类结果，这类方法识别工艺相对复杂，对非专业人员并不友好。借助木材切面的图像信息或光谱信息可以简单方便地对木材进行分类，然而由于不同种木材之间存在的近似性，这类方法往往分类精度不高或只适用于某些阔叶木材。提出了一种基于木材横切面图像信息和光谱信息的多特征木材分类算法，首先分别采集木材横切面的光谱信息以及图像信息；再使用Segnet图像分割方法将待分类样本分成含管孔木材和不含管孔木材两组，并对含管孔样本组中的木材进行管孔分割；然后对含管孔样本组中的木材提取管孔特征、光谱特征以及纹理特征，对无管孔样本组木材提取光谱特征和纹理特征；最后根据这些特征使用支持向量机分别对木材进行分类并记录其木材的分类结果，对分类结果不一致的样本使用相似性判据判断最佳分类结果。为了验证该方法的有效性，以20种常见的阔叶木材和针叶木材的混合样本集为研究对象，对其进行了分类。实验结果显示三种特征均可以对木材进行分类，单独使用光谱特征、纹理特征以及管孔特征对木材进行分类的最高正确率分别为93.00%，89.33% 和69.23%，通过相似测度的判断后三个特征可以相互补充从而进一步提高木材的分类正确率，最高正确率可达98.00%。综上所述，该方法可以对包含阔叶木材和针叶木材的混合样本集中的木材进行分类，木材横切面的光谱特征、纹理特征以及管孔特征可以相互补充，从而使分类正确率进一步的提高。与目前的主流木材分类方法进行对比，发现该算法的分类正确率高于其他算法。

Study on Simultaneous Classification of Hardwood and Softwood Species Based on Spectral and Image Characteristics

Wood is an indispensable renewable resource in people’s lives, and it also plays a vital role in architecture, craft, furniture, structural material and so on. The common wood species in the market are various, and the quality and price of different wood species also differ very much.Therefore, the use of intelligent technology to undertake correct wood classification can prevent illegal trader’s shoddy product and reduce the workload of wood classification personnel greatly. Though accurate wood classification results can be obtained through the genetic and anatomical information of the wood sample, the identification process of these two methods is relatively complex, not easy for non-professionals. With the help of image information or spectral information of wood surface, wood species can be classified and conveniently. However, due to the similarity among different wood species, the classification accuracy of these two methods is often not high or only suitable for some specific wood species. Therefore, we propose a multi-feature wood classification algorithm based on the image information and spectral information of wood cross-section. First, spectral reflectance curve and image information of wood cross-section are collected, respectively. Then, the Segnet image segmentation method is used to divide the wood samples into two groups: wood with and without pores. The characteristics of pores, spectral features and textural features are extracted from wood species with pores, and the textural features and spectral features are extracted from wood species without pores. Next, according to these characteristics, a support vector machine (SVM) is used to classify wood and record the classification results. Finally, the similarity criterion is used to judge the best classification results for the samples with inconsistent classification results. In order to verify the effectiveness of the method described in this paper, the mixed sample set of 20 common hardwood and softwood species is used and classified. Experimental results show that these three wood features can be used for classification, and the highest wood recognition rate is 93.00%, 89.33% and 69.23% for spectral, textural and pore features, respectively. By similarity measurement, the three wood features can complement each other so as to improve further the wood species classification accuracy with the highest recognition accuracy of 98%. To sum up, the method described in this paper can be used to classify a mixed wood sample set that includes hardwood and softwood. The spectral features, textural features and pore features of the wood cross-section can complement each other, thus improving classification accuracy. In addition, in this paper,we also compareour method with the state-of-the-art wood species identification methods and find that the classification rate of this algorithm is higher than other algorithms.