| In recent years,the technology of panel defect identification based on convolutional neural network has gradually become a hot research direction in machine vision technology,which has great significance for improving the production efficiency of wood processing enterprises and the use value of wood boards.In view of the problems of omissions and mischecks in traditional manual defect detection,this thesis takes common defects of wood boards such as cracks,holes in wood boards and scarring of trees as the research object,and makes an in-depth study on the defect identification of wood panels based on Yolov3 network in the wood processing assembly line.The main contents of this thesis are as follows:The image preprocessing is carried out on the defect image of wood board to collect images of normalized processing,enables the image data sets on the category to uniform distribution,and then USES the data amplification method to increase the number of wood defect images data set,make the wood hole image number of 6243 copies,wood crack image number of 9700 copies,tree scar image number of 7534 copies,finally using professional software making VOC data sets for defect areas of wood boards;In this thesis,through image preprocessing such as geometric transformation method and color space transformation method,the number of images is increased,the quality of images is improved,and the problem that the training samples will be seriously biased to the category of a large number of samples and the problem of insufficient generalization ability on the test data set is successfully solved.In order to improve the simplicity and efficiency of the network model in the defect detection of wood boards,this experiment mainly uses the threshold segmentation and morphological processing methods to remove impurities such as chip holes in the pictures of wood boards.The speed of defect detection of wood boards reaches 0.1534,and the accuracy reaches 93.34% for wood defect images information loss,wood defect problems such as leak,in this thesis,the image processing algorithm and Yolov3 network integration of multi-scale feature detection,through the Yolov3 network model parameter setting and optimization,image feature extraction,and the test results of three dimension characteristics of wood defects,it is concluded that the input sampling 32 times under the wood defect images suitable for detecting large scale images,such as cracks in wooden boards;Subsampling 16 times is suitable for detecting medium scale images,such as tree scarring;The subsampling of 8 times is suitable for detecting small-scale images,such as wood holes,making the m AP of wood holes reach91.08%,wood cracks 98.25% and tree scars 90.84%,which greatly improves the detection speed and accuracy of Yolov3 network and reduces the probability of missed detection.Compared with Faster R-CNN network and SSD network,the MAP of Faster R-CNN network was 92.78%,the detection speed was 0.6172 s,and the MAP of SSD network was76.89%,and the detection speed was 0.1665 s The m AP of Yolov3 is 93.34%,and the detection speed is 0.1534 s.Therefore,it can be concluded that the algorithm used in this thesis has a higher detection accuracy and a faster detection speed.This conclusion is well verified by comparing the board detection renderings of the three algorithmsBased on the above three points,this thesis conducts comparative experiments and the results show that the Yolov3 network model algorithm has shorter training time,lower miss rate,more accurate positioning,and Faster detection speed.Yolov3 network model is superior to Flyer in board defect detection Faster R-CNN network and SSD network can meet the demand of real-time detection of the assembly line of wood production and processing enterprises. |
