Research On Recognition Method Of Wood Surface Defects Based On Machine Vision

The detection of surface defects in wood is crucial for the evaluation of the wood material,which in turn determines the use and economic value of the wood product.If the type and location of surface defects can be detected precisely and quickly,and then separated and removed in a rational way,the utilization of the wood can be increased while ensuring the quality of the wood product.In this context,it is necessary to combine the wood industry with modern technology,so that automatic,efficient and intelligent machines can be used in wood production.This paper takes wood surface defect images as the research object,and accomplishes automatic classification of wood surface defects by digital image processing and artificial intelligence technology.This paper carries out research on image pre-processing,feature extraction,feature dimensionality reduction and model building to realize the recognition of wood surface defects based on machine vision.The main research contents are as follows:(1)In terms of data pre-processing,this paper performs data enhancement on the original data from four perspectives of geometry,noise,brightness and filtering,which not only achieves data expansion but also improves the visual effect of the image and highlights the differences between features in the image.In the feature extraction part,based on the four aspects of gray statistics,color,shape and texture,this paper calculates the five features of wood image,including gray statistics,color moments,Hu invariant moment,local binary pattern and gray gradient co-occurrence matrix,and then combines these five features into the feature expression set of the original image.(2)For the problems of using multidimensional features to characterize images,which may generate redundant information and lead to an increase in computation and affect the recognition effect of the model,this paper uses three dimensionality reduction algorithms,linear discriminant analysis,principal component analysis and isometric feature mapping,to parsimoniously optimize the extracted features for the subsequent model analysis.(3)Based on the principles of three models in machine learning:random forest,gradient boosting decision tree and support vector machine,the corresponding classification models were built to complete the identification of wood defect types.In the comparison experiments,the data after dimensionality reduction is input to different classification models for experiments,and the performance indexes of the models are compared to select the optimal combination of dimensionality reduction method and classification model.According to the experimental results,the linear discriminant analysis combined with support vector machine is the best,and its precision rate,recall rate,F1-score and accuracy are 89.35%,88.11%,88.28%and 88.55%respectively.(4)In the deep learning-based research,three convolutional neural network frameworks,AlexNet,VGGNet and ResNeXt,were selected in this paper,and the network model structure and parameters were designed and adjusted to complete the recognition of wood surface defects.According to the experimental results,the three models can effectively complete the recognition task of wood surface defects.Among them,VGGNet model has the best performance index,and its precision rate,recall rate,F₁-score and accuracy rate are 97.33%,97.01%,97.1%and 97.34%respectively.