Research On Surface Defect Detection System Of MDF Based On Deep Learning

MDF is the main product of wood-based panel industry in China.After more than 20 years of development,it has formed a relatively complete production process system.However,in the current MDF manufacturing process,the detection of surface defects is still mostly dependent on human.The manual method is slow and inefficient.Besides,due to the different understanding of the inspectors and the variety of the surface defects,it is difficult to give an objective and accurate evaluation of the surface quality of MDF.Therefore,it is very important to find an accurate and efficient method to detect the surface defects of MDF for wood products enterprises to improve product quality and reduce production costs.Aiming at 6 kinds of defects of MDF,such as oil,soft,black spots and sundries,this paper designs a surface defect detection system based on deep learning.The system uses industrial linear camera to collect the surface image of moving plate.The image is input into the deep learning algorithm after preprocessing such as noise reduction.The detection and quality grading of defect type,defect location and defect area size are carried out.The detection results are displayed in the upper computer software in real time and saved in the local database.In this paper,the current mainstream deep learning instance segmentation algorithms Mask R-CNN,Yolact,Centermask were used for defect detection experiments,and the feature extraction network and image segmentation network of Centermask algorithm were improved.The algorithm evaluation experiment and the whole system test experiment were carried out on the workstation of GPU P2000.The experimental results of algorithm evaluation show that the improved Centermask algorithm has54.5%and 46.6%defect recognition accuracy index m AP_50-95and defect segmentation accuracy index m AP_50-95 respectively,which is 0.3%and 1.2%higher than Centermask algorithm.The overall test results show that the system’s defect detection performance index F1-score is 79.0%when the detection speed is17 sheets/min.