Wood Surface Defect Detection Based On Deep Learning

Wood surface defect detection is an important part of wood processing and intelligent wood processing industry.In order to improve the utilization efficiency of wood,scholars all over the world have proposed a variety of wood surface defect detection methods.However,the texture structure,color and size of wood surface defects are obviously different,which makes it difficult to locate and identify the defects.First of all,there is no public data set of wood surface defects in the current wood defect detection,which can not achieve the unified detection and recognition of the defect area.Then there is no easy algorithm to detect defects.In this regard,this paper mainly from two aspects of defect detection and defect recognition.The research contents include the following aspects:(1)In this paper,five types of defects including knots,holes,decay,cracks and browning are collected and six types of wood surface images are collected from normal wood surface images.Use cutting,affine transformation,rotation,mirror flip,random noise,image enhancement and other methods for data enhancement.(2)Built a wood surface defect detection model based on migration learning and deep learning.First,use the pre-trained weights on the PASCAL VOC and MS COCO data sets to initialize the Faster R-CNN,YOLOv3 and SSD models to train the wood surface defect detection model in this article.Compare the three models with the same hyperparameter settings and find that YOLOv3 and SSD are more suitable in this data set,the two models are optimized through three aspects: confidence setting,K-means frame prediction,and feature extractor lightweight.The results show that the optimized model can detect and locate wood surface defects better.(3)A recognition model of wood surface defects based on deep transfer learning was built.Compare and analyze the classic deep learning models Alex Net,Inception-v3 and Res Net50,transfer learning Image Net large data set pre-training model to perform image classification and recognition,select the appropriate classifier to improve the model accuracy,and finally multi-model fusion enables the detection of small samples of wood defect data sets has a higher accuracy rate.The experimental results show that the model built in this paper has high detection accuracy in the data set,and has a good ability to locate and classify wood defects.In the actual wood processing,the efficiency of the use of wood can be improved,and it has certain industrial use value.