Micro-spectral Wood Classification Based On Fusion Of Spectral Information And Texture Information

Woods are natural organic polymeric compounds,which are significant renewable raw materials for the national economy and essential living resources.However,the quality of wood in the market is uneven,and many businesses deceive consumers with sub-optimal woods,because many wood textures are similar and difficult to distinguish based on the naked eye.There are many difficulties in dividing the wood texture and identifying the authenticity of the wood.Therefore,the classification and identification methods for wood identification still need to be updated.At present,the classification algorithm is still inadequate,and most classification processes rely on the spectral information of the image,while the depth of the wood texture informations is not intensive enough.In addition,the support vector machine in the supervision classification algorithm has its unique advantages in dealing with nonlinear problems and dealing with dimensional disasters,but the selection and application of kernel functions are relatively simple.Under the framework of complex kernel function,an effective and syncretic complex kernel function method of spectral information and wood texture information was proposed.In terms of the microscopic hyperspectral data informations of twenty wood species,the main contents and work of this article are as follows:(1)Image Texture Feature Extraction of Microscopic Hyperspectral Images Based on Fractal TheoryBy introducing fractal theory into the wood classification of microscopic hyperspectral images,this paper proposed the classification of microscopic hyperspectral wood texture information using multiple fractal spectra.We reduced the number of bands by using adaptive band selection and K-L divergence to select characteristic wavelengths,and then proposed three different density functions:Image Brightness Function,Image Gradient Function,and Image Laplacian Function.The texture images corresponding to the selected bands were divided into several binary images,and the box dimension method was used to solve the fractal dimension of the divided image to obtain multiple fractal dimension values,which also corresponded to multiple fractal curves,and finally the multifractal curves were obtained by averaging.Compared with traditional texture extraction algorithms,such as gray level co-occurrence matrix and wavelet texture extraction,the proposed multifractal method has greatly improved recognition accuracy.(2)Classification of wood species based on fusion of spectral information and image texture informationWe selected the region of interest with a certain size for the microscopic hyperspectral data,extracted the average spectral data,and divided the data sample set.Three spectral preprocessing algorithms—Savitzky-Golary convolution smoothing algorithms,multivariate scattering corrections,and standard normal variable changes,were used to reduce noise and smooth it.Then,two algorithms,competitive adaptive reweighting algorithm and continuous projection algorithm,were used to reduce the dimensionality of the spectrum,select the characteristic wavelength,and extract the spectral information of 20 kinds of wood.A single classification feature cannot fully be expressed in terms of the unifying characteristics of microscopic hyperspectral data.This paper fused spatial texture information and spectral information,and sent it to the classifier.This paper used many supervision classification methods—support vector machine,correlation vector machine and BP neural network.Finally,the experimental comparison showed that the support vector machine had better classification effect.Support vector machine had a good classification ability when dealing with samples with limited labels.Therefore,within the framework of composite kernel functions,five composite kernel functions were proposed to establish a graph fusion classification model.By comparing with the single feature information model,the second complex kernel function is better in recognition effect,with a recognition rate of 97.26%.Therefore,the methods,using composite kernel function support vector machine to fuse the image information,could improve the complementary utilization rate between different features and increase the classification accuracy.