Research On Wood Classification And Sorting Algorithms Based On Image Multi-Feature Pattern Recognition

China's forestry resources are relatively scarce,but the current domestic mainstream wood processing plant processing methods are too simple and extensive,traditional mode,resulting in low utilization rate of raw materials processing,processing efficiency is poor.In order to effectively improve wood utilization and wood production efficiency,this paper classifies wood based on multi-feature pattern recognition.By searching and classifying the selected wood images according to tree species,wood classification based on machine vision is further realized according to defect categories and inspection standards.The existing wood defect classification based on machine vision is to classify the wood grade according to the morphological characteristic parameters of the defect itself.However,it has not been achieved on the basis of classification of wood species.This is inconsistent with the classification operation of wood grades in the actual production process.In view of the above problems,this study classifies wood image defects based on the discrimination of wood species.According to the different types of defects between coniferous wood and broad-leaved wood,and according to the extracted defect area measurement parameters,the wood is classified into special,first,second and third grades.In this study,50 representative Chinese common tree species were selected from 120 tree species.Each tree species had two images corresponding to the diametral and chord sections.As the research object of the retrieval experiment,a multi-species sample bank was constructed.Samples of 5 common tree species in Northeast China were collected to make wood images of two kinds of cut surfaces(diameter cutting and chord cutting)of these five tree species,100 of each tree species,totaling 1000 pieces,including 104 dead knots,40 live knots,72 insect pests and 92 crack defects.Sample bank 2 was constructed.The main research contents and experimental results are as follows:(1)In order to realize wood species retrieval and discrimination,this paper studies the methods of color and texture feature extraction and similarity matching and discrimination in wood image retrieval.According to the characteristics of wood color and texture,a matching retrieval method suitable for wood species identification was proposed.The main tone feature extraction method based on unequal interval quantization in color space can be used to match wood images with narrow color distribution and relatively subtle color differences among tree species,so as to achieve a better retrieval and discrimination effect.After color feature retrieval,six texture features including contrast(CON),second-order angular moments(ASM),variance sum(SV),long-range aggravating factor(LRE),fractal dimension(FD),proportion of horizontal energy distribution of wavelet transform(EPLH)are further used to construct feature quantities system for further retrieval.The experimental results show that step-by-step retrieval can make the retrieval discriminant results tend to be higher accuracy.(2)A binary local threshold segmentation algorithm for wood image defect is proposed,which calculates the threshold of each point by calculating the mean,standard deviation and extremum of the window template.The experimental results show that the proposed algorithm has good performance for wood image defect segmentation in complex background,and its performance is obviously superior to the global threshold,and Bernsen algorithm,and slightly higher than Niblack algorithm and Savola algorithm.The accuracy of defect segmentation of wood image can reach 92.58%,which is more suitable for wood defect images with uneven color or illumination and texture noise interference.(3)On the basis of defect segmentation,according to the specificity of different wood defects,morphological methods are used to further improve the defect extraction effect,and on the basis of greatly reducing noise,the defect binary features are restored to the greatest extent.The length,width,number and location of defects are extracted by morphological function.(4)Three models are used in wood image defect detection and classification:BP neural network model,SVM support vector machine classification model and CNN convolution neural network model.BP neural network achieves classification by extracting LBP features and HOG-LBP fusion features respectively,with the highest classification accuracy of 50%.SVM classification model extracts HOG feature,LBP feature and HOG-LBP fusion three groups of features,and the defect classification effect using HOG-LBP fusion feature is significantly better than using HOG or LBP single feature.Four kinds of kernels are used to realize classification based on HOG-LBP fusion features.Experiments show that polynomial kernels and Gauss kernels have the best classification performance,with an accuracy of 98.68%.The dimension of input layer of CNN convolution neural network model is 512 x512 x3,and the accuracy of classification is verified when the number of layers is 2-4 by using convolution kernels of 1 and 0 phases of 9 x9.CNN convolution neural network is used to test the optimal four-layer structure,and the classification accuracy is 98.68%.(5)By optimizing the experimental parameters,monitoring the training process and comparing and analyzing the experimental results,the performances and advantages of different models are verified.Through comparative analysis,it is concluded that BP neural network is a classical model,but it is not sensitive to the HOG,LBP and HOG-LBP fusion features of wood defects.Therefore,it is not suitable for classification of wood models under the current situation of feature extraction.CNN convolution neural network model and SVM support vector machine model are more suitable for wood defect detection and classification,and they have higher classification accuracy for wood defect detection and classification.