Study On On-line Detection System For Surface Defects Of Particleboard Based On Machine Vision

The continuous press production line is the most advanced particleboard production equipment at present.However,due to raw materials,production technology and other reasons,there may be some defects in the surface of products including big shavings,glue spots,oil stains,soft defects and imperfect sanding defects,etc.Surface defects can reduce the strength of the panels,bring difficulties to the secondary processing and cause great economic losses to the manufacturers.However,at present,domestic manufacturers still rely on workers to use naked eyes to detect board defects.The workers observe the panels for a long time which cause visual fatigue and high rate of missed detection and false detection.At present,the researches on the surface defect detection in China and abroad mainly focuses on the wood,rotary cut veneer and plywood.There are few researches on particleboard,and there is no mature automatic particleboard surface defect detection system put into the production line.Therefore,the development of automated particleboard surface defect online detection system has become an urgent need in China's particleboard production industry.In this paper,the machine vision technology was applied to the surface defect detection of particleboard,and the online defect detection algorithm was designed and developed.The detection algorithm realized the automatic acquisition of detection areas,correction of the panel images,fast location of defect areas,segmentation of the defects and the identification of the defect types.The system was built on a continuous press production line,and the online detection of five common defects of big shavings,glue spots,oil stains,soft defects and imperfect sanding defects was realized.The main research contents and conclutions are as follows:(1)The hardware platform of the on-line detection system for particle board surface defects was designed and built.The hardware equipments were selected,and the control flow of the system was designed and implemented.(2)Automatic acquisition and image correction of the detection area were realized.A gradient adaptive threshold edge detection algorithm was proposed.By calculating the gradient matrix of the image and using two proportional numbers to adaptively determine the gradient thresholds of the edge,the problem that the edge detection result was affected by the image brightness was solved.This algorithm was combined with the Hough transform to calculate the tilt angle of the board surface so as to realize tilt correction and automatic acquisition of detection area.Aiming at the problem of uneven illumination of the image,a correction method based on gamma transformation and image difference was proposed to correct the gray value while maintaining the difference of the gray value between the normal board surface and the defect areas.(3)The rapid detection of board defects was achieved.A fast localization method for defect regions based on gray mean value classifier and variance classifier was proposed.The gray value matrix and variance matrix of the image were established and regional connectivity was done to locate the defect regions.An adaptive fast mul-threshold image segmentation algorithm was proposed to automatically determine the number of segmentation thresholds and the threshold search strategy was optimized to achieve fast segmentation of defect regions.Using the detection algorithm of this paper,600 images collected from the production line were detected.The normal board misdetection rate was 5%.The big shavings and imperfect sanding defects were all detected.The miss detection rate of glue spots,oil stains and soft defects reached 1.6 %,4.0%,and 6.6%.The average detection time of one image was 867 ms.(4)A random forest classifier and a lightweight convolutional neural network with 8-layer were designed and implemented to classify the defects.The average accuracy and recall rate of random forest reached 94% and 95% respectively.The total time of the feature extraction and prediction for each sample was 41.28 ms.The average accuracy and recall rate of Convolutional Neural Network(CNN)reached 87% and 89%,and the predicted time of one image was 5.4ms.The CNN has a higher speed but the classification effect of CNN is not as good as random forests,but this provides a new method for identifying the types of board defects in the big data environment in the future.(5)The system was implemented in the continuous press production line of the particle board.The average detection time of each panel surface was 1922 ms,the false detection rate was 5.1%,and the missed detection rate was 2.7%.The accuracy of defect detection reached 97.3%,which met the online detection requirements of continuous press production line in terms of time and effect.Compared with the accuracy rate of similar systems abroad of less than 70%,the detection effect has been greatly improved.On-line detection system for particleboard surface defects based on machine vision realizes automatic detection of surface defects of particleboard in continuous press production line and reduces the missing detection rate and false detection rate of defective boards.And the economic losses caused by the board defects of enterprises also will be reduced.The system can improve the automation level and operation efficiency of the production line and the modernization development of the particleboard industry in China will be promoted.