Research On Nondestructive Testing Method Of Solid Wood Board Based On Deep Learning

As people’s requirements for the quality of building,home,handicrafts and other necessities of life continue to improve,the wood market competition intensifies,the wood market for wood processing industry product quality requirements are increasingly stringent.However,the domestic wood processing industry relies too much on manual processing,which leads to uneven product quality and low utilization rate of wood resources.Therefore,it is particularly important to develop solid wood board intelligent processing equipment,transform and upgrade the wood processing industry to intelligent manufacturing,and realize automation and intelligence.The non-destructive testing task of solid wood board will play a decisive role in the intelligent processing system of solid wood board because it can replace manual work to obtain the image feature information of the board surface and make decisions.Therefore,this paper combines machine vision technology and deep learning technology to study the nondestructive testing method of solid wood board.The research is as follows:(1)An online image acquisition system for solid wood board was established.The board images of six varieties of Chinese fir board,beech board,cherry board,ash board,birch board and pine board were collected successively.After preprocessing operations such as background removal and filtering denoising,the defect detection data set of solid wood board and tree species classification data set of solid wood board were produced respectively.(2)In view of the low accuracy of defect location and classification in the application of traditional SSD algorithm in the surface defect detection of solid wood,Dense Net121 network is introduced.The idea of residual learning is used to avoid the loss of information in the feature map and increase the depth of the network so as to obtain deeper information.The experimental verification shows that compared with SSD algorithm,Faster RCNN and YOLO algorithm,the average accuracy is increased to 96.1 %.The P-R curve corresponding to various defects is steeper,and the area AP value under the curve is closer to 1.(3)Aiming at the problem of low classification accuracy of traditional image feature extraction method combined with classical machine learning for tree species classification of solid wood,attention mechanism and spatial pyramid pooling strategy are introduced into Res Net101 network.The obtained AM-SPPRes Net convolution layer has better feature extraction ability than VGG16,Res Net50 and Res Net101.The support vector machine is used to replace the full connection layer to learn the feature information obtained by AM-SPPRes Net,which greatly improves the classification accuracy of tree species of solid wood,and the classification accuracy on the test set reaches 99 %.Compared with other models,it has more balanced prediction performance.(4)A nondestructive testing system for solid wood board was developed.A set of non-destructive testing system software for solid wood board was developed by using Qt software,Open CV and Libtorch.The system includes main functions such as image acquisition,defect detection,tree species classification,results display,data query and son on,and has good human-computer interaction.