Recognition Of Wood Boards Surface Defects Based On Object Location And Near Infrared Spectrum Technology

The surface defects of wood will affect the quality and grade of wood board.Near infrared spectroscopy can be used for qualitative and quantitative analysis of the detected objects through the information contained in its spectra.Because of the difference of physical shape and the composition and content of organic matter,the absorption spectrum in the defect area is different.Therefore,the near infrared spectroscopy can be used to identify the defect area.In order to improve the recognition speed and accuracy of the solid wood surface defect,a method of nondestructive detection of wood defects based on near infrared combined with object location of machine vision is proposed in this paper.The study takes 450mm x 150mm x 25mm of larch wood boards as the experimental samples,designs a defect locating method and selects the spectral processing method which is suitable for wood defects.Through extracting the effective information in the spectra,it builds the recognition model of wood surface defects.The specific research results are as follows:JIn order to solve the problem that the surface defect of the solid wood plate is not complete and the speed is slow,the FDBC-Grabcut algorithm is used to locate the defects quickly.The algorithm improves the traditional Grabcut algorithm from two aspects of fast convergence and fractal theory.First of all,this method reduces the resolution of the original image,Iterates the Grabcut algorithm in the reducing image,and the convergence speed is accelerated greatly;in addition,the differential box counting algorithm is applied to detect the outlines of the defects,which solves the artificial interaction problems of the traditional Grabcut.The experimental results show that the FDBC-Grabcut algorithm can fast and fully locate the wood surface defects in the images that are processed by HSV color space transform and median filter.The average location time is 0.561s.To avoid the influence on the recognition performance of the model with non-unif-orm due to random division of sample set an improved K-S data set division method is proposed to divide the wood defect samples into training set and test set,the ratio is 3:1.In the improved K-S algorithm,the original spectra are reduced by PCA and the computational complexity is reduced.Then the normalized Euclidean distance is used to replace the traditional Euclidean distance to improve the computational efficiency of the algorithm.In addition,in order to solve the problem of spectral baseline drift and high frequency noise in the process of spectral acquisition,the spectra are preprocessed to make the spectral profile more clear.The experimental results show that using the method of first derivative combination of SG smoothing to preprocess the sets can provide the best preprocessed effects with the highest Rcv2 of 0.8220 in the principal component regression model,and the RMSECV of 1.7416 is the minimum.Since the high dimension and large redundancy of the spectra that are easily to affect the modeling speed and precision,the use of sparse dimension(Sparse Reduction,SRE)method for feature extraction of the spectra is proposed.SRE uses the linear measurement of sparse coefficients to complete the selection of the nearest neighbor values from the sparse data,which reduces the dimension of the original spectral data.The experimental results show that compared with other feature extraction methods,using the spectral features extracted by the SRE algorithm proposed in this study to establish defect recognition model greatly improves the model recognition performance.The established discriminant partial least squares(DPLS)defect recognition model with the spectral features extracted by the SRE algorithm gets the highest Rcv2 of 0.9016,and lowest RMSECV of 1.2952.The a-verage recognition rate of defects is 90.4%.This study proposes a CPSO-OMP compressed sensing classification model for the recognition of defects on the surface of the wood,this model recognizes the defects accurately according to the near infrared spectral features of defects.To further reduce the high dimensional inner product operation of the OMP iterative matching algorithm,the particle swarm optimization(PSO)algorithm is used to find the extremum quickly;for existence of premature stagnation problems,the chaotic sequence(Chaotic sequence)is applied to avoid the premature of PSO,make PSO jump out of local minima and achieve the global optimization.The experimental results show that compared with DPLS,BP neural network and LS-SVM classifier model,the CPSO-OMP compressed sensing classifier proposed by this study is capable of the best average recognition accuracy for surface defects of wood with the highest recognition rate of 97.6%...