Research On Defect Detection Algorithm Of Solid Wood Board Based On Extreme Learning Machine

With the development of the wood processing industry,the demand for the quality of wood processing and the utilization of wood is becoming more and more high.The traditional methods of wood detection have not met the demand.It is very important to improve the technology of wood detection and classification according to different needs.In recent years,wood nondestructive testing technology has grown up and widely applied in wood processing industry.This paper takes solid wood board as an example to collect three kinds of defects,such as knots,holes and cracks on the surface of solid wood.Image processing technology and classification recognition technology based on extreme learning machine has been combined to identify and classify solid wood board.The detection and recognition of solid wood surface defects included the comprehensive application of machine vision and pattern recognition.Five aspects of the detection system were described in this paper,including the composition of the detection system,sample selection and image acquisition,image segmentation,feature extraction,recognition and classification.The main research work was as follows:First,a nondestructive testing system for solid wood board was built.It mainly included two parts: the mechanical structure part and the image acquisition system.The mechanical structure was composed of conveyor belt,mounting device of scanner,control and signal induction device.It was responsible for the transportation of wood and the start and stop of the acquisition system.The image acquisition system included two Chroma+Scan3350 laser profile scanners,which integrated two dimensional plane scanning and three dimensional contour scanning,and was equipped with LED lighting source.A total of 400 images of Pinus densiflora and Pinus sylvestris lumber were collected.Secondly,the image was pre-processed by grayscale,gray enhancement and filtering denoising.The interference factors in the acquisition process were eliminated and data was simplified.The threshold segmentation methods based on Otsu,maximum entropy method,and histogram segmentation and the edge detection algorithm based on Roberts operator,Sobel operator,LOG operator and Canny operator were compared.Finally,image segmentation was done by combining histogram segmentation and edge detection based on Sobel operator.Then the defect area was extracted after mathematical morphology transformation.Thirdly,Feature extraction methods based on grayscale and geometric features were used to extract feature vectors.Three gray features,six geometric features and seven moment features were extracted.Principal component analysis was applied to reduce the 16 dimensional vector to 6 dimensional vector.These feature vectors were used as classifier inputs to classify defects.Finally,a classifier based on extreme learning algorithm was designed,which was improved by combining the AdaBoost algorithm.By comparing the classification results,the limitations of single classifier and the advantages of improved classifier were analyzed.The experimental results show that: For a single extreme learning machine classifier,the algorithm was simplified though random initialization of input weights and hidden layer bias.Compared with the traditional classification algorithm,it had a certain degree of improvement in learning speed and prediction accuracy.But its randomness would affect the accuracy and stability of the classification results.The improvement of extreme learning machine itself was also limited.Therefore,the classifier improved by AdaBoost algorithm was developed to further improve the prediction accuracy and stability of the algorithm.