Search

The image information and spectral information of wood sections can be used to identify wood species. Hyperspectral images have both image information and spectral information, but they have disadvantages such as large data capacity, slow reading speed, and the necessity of expensive equipment for their acquisition. In this study, the classification results of Pterocarpus by using visible/near infrared (VIS/NIR) spectral information and RGB images were compared with hyperspectral images. The VIS/NIR spectral curves, Hyperspectral, and RGB images of five wood species of Pterocarpus with similar transverse-sections were collected. In feature-level fusion, the feature vectors are directly connected in series, and features fused by canonical correlation analysis are compared. In decision-level fusion, an extreme learning machine and a composite-kernel support vector machine (SVM) are used and compared. In the feature- and decision-level fusion methods, the recognition results of VIS/NIR spectral curves plus RGB images were largely similar to those of hyperspectral images. Therefore, a recognition effect similar to that of the hyperspectral image can be obtained by collecting the spectral information and image information of wood sections separately, which can reduce the cost of data acquisition and improve the speed of data processing. [ABSTRACT FROM AUTHOR]

In this study, a method based on fuzzy gray level aura matrix (FGLAM) textural feature and spectral feature fusion is proposed to improve the accuracy of wood species classification. The experimental dataset is acquired by two sensors. The visible images of wood samples are acquired using a RGB image acquisition platform, and the spectral curves are acquired using a USB2000-VIS-NIR spectrometer having a spectral band of 350–1100 nm. Firstly, the textural features are computed by using FGLAM for gray level wood sample images. The original spectral curve is smoothed and corrected to remove noise. Secondly, the obtained textural features and spectral features are processed in extreme learning machine to complete the decision-level fusion and obtain the classification results. The experimental results for 15 wood species show that the classification accuracy of the textural feature by using FGLAM alone can reach to 90%, which is better than that acquired using the traditional gray level co-occurrence matrix, local binary pattern, and the mainstream improved basic gray level aura matrix. Moreover, the classification accuracy based on the FGLAM textural feature and spectral feature fusion proposed in this paper can reach to 100%. Finally, for 8 similar wood species, the recognition accuracy of the present fusion scheme reaches to 98.75%, which is also better than those state-of-the-art methods. [ABSTRACT FROM AUTHOR]