Research On Classification Method Of Tree Species Based On The Spectrum And Texture Features Of Wood Cross-section

Wood is integral to people’s lives,due to its important applications as structural material,fuel,decoration,and other uses.The classification of wood is a critical part of determining how it is used.The traditional manual classification method is not only inefficient,but also prone to errors.In contrast,classification of wood by a computer is highly automated and can also give better results,which has important practical significance.In this study,accurate classification of wood was achieved by utilizing cross-sections with rich spectral and image information.Hyperspectral and RGB images,as well as one-dimensional visible/near-infrared(1-D VIS/NIR)spectral information,were collected for the cross-sections of 50 types of commonly used wood.From this comprehensive dataset,the following aspects were studied in this paper:(1)The spectral features of wood cross section different structure are not consistent,and the use of average spectrum may affect the wood classification effect to a certain extent.To solve this question,the spectral reflectance corresponding to the pore center and periphery(axial parenchyma)on the wood cross-section was extracted by morphological and spectral clustering analyses.The spectral reflectance separability of pore center,pore periphery,and random selection were also discussed.These spectra were processed by dimensionality reduction and other pretreatment methods.The thus-treated spectra were sent into classifiers for training,and then unknown wood samples were classified using the trained classifiers.The spectral reflectance separability of the pore periphery was found to be greater than that of the pore center,which resulted in improved classification when using this spectral information.(2)Outside noise interferes with wood classification.To solve this question,a wood classification method based on the fusion of spectral features and texture features was proposed.First,several bands were selected from the hyperspectral image by the band selection method based on similarity,and the images of these bands were fused into a gray image by the image fusion method based on wavelet transform.Then,texture features were extracted from this image based on the PLS and LBP theories.Then,the near-infrared spectral reflectance and texture features are connected in series to generate feature vectors,which were sent into the classifier to realize the classification of wood species.It was found that the texture features and spectral features were not sensitive to changes in light;hence,this experiment proved that this method can be used to effectively classify wood species even when there is hyperspectral image distortion.(3)The amount of hyperspectral image data generated is very large,and so is the cost of data collection.In this study,the RGB image of wood cross-section and 1-D VIS/NIR spectral information are considered the research objects;the fractal principle and the LBP theory are used to extract spectral features of 1-D VIS/NIR and texture features of RGB image from a wood cross-section.These two features were fused by the feature-level fusion method,based on the results of typicality correlation analysis,and then used for classification.The experimental results showed that the classification accuracy of the feature fusion method was close to that of hyperspectral images,and the classification speed was much higher.(4)we discuss the characteristics of the three methods used to describe and classify wood:hyperspectral imaging of wood cross-sections,1-D VIS/NIR spectroscopy,and RGB imaging.Experimental verification of the wood classification method described herein and the existing ones revealed that the spectral characteristics of the hyperspectral images are stronger than those of 1-D VIS/NIR spectra,while the texture characteristics of the hyperspectral images are slightly stronger than those of RGB images.To compare the effects of the three data sets on the classification of tree species,the spectral and texture properties of hyperspectral images were fused at both the decision level and the feature level after information fusion of 1-D VIS/NIR spectra and RGB images.However,the experimental results show that the classification accuracy of the fusion of the 1-D VIS/NIR spectrum and the RGB image approached that of the hyperspectral image;therefore,the 1-D VIS/NIR spectrum and RGB image can be used to replace the hyperspectral image when faster classification is required.Lastly,this paper summarized existing wood classification methods.To summarize,this study addressed how computer classification methods based on image and spectral information can be applied for classifying wood.The characteristics of the different methods were examined,as well as how they handled the complex and real-world situations of identifying wood cross-sections.The methods explored in this study will greatly assist in professional wood identification.Beyond this,these methods can be generally applied to solve other classification problems.