Study On Wood Surface Defect Indentification Based On Digital Image

Furniture industry is the traditional advantageous industry of Nankang,the most characteristic and largest industrial cluster in Ganzhou at present,and also one of the exemplary model of industrial clusters focused on Jiangxi Province.As the raw material of furniture,the quality of wood is the prerequisite to ensure the performance of furniture industry.Digital image-based wood surface defect detection is an important hot issue in the wood industry research field,and its superior detection effect greatly alleviates the problem of low efficiency and low accuracy of manual detection.Therefore,the research of digital image-based wood surface defect detection is of great significance to the wood industry production and even national economic growth.This paper first introduces the theory related to digital image-based wood surface defect detection,and then carries out researches using traditional and deep learning methods,including:(1)For traditional digital image features are not easy to design manually,this paper proposes a new feature extraction method based on HOG and GLCM features which is called HOG-KPCAGLCM.Firstly,the KPCA is used to reduce the dimensionality of the obtained HOG features,and then the influence of different dimensions on wood surface defect detection is experimentally studied to obtain the optimal HOG-KPCA feature dimensionality.Secondly,the HOG-KPCA feature is cascaded with GLCM and the effect of different weights is experimentally studied to get the optimal weights.Based on the experimental wood surface defect dataset,SVM is used as the classifier.The experimental results show that classification accuracy of the proposed method in this paper is 85.05%on the test set,which is better than those of HOG(80.67%)and GLCM(68.62%).(2)Aiming at the problems that the training accuracy of deep learning network is not excellent enough and the training time is long,this paper proposes an improved Resnet-50 network based on CSP and attention mechanism and Ranger optimization method is introduced to train the network after studying the gradient transfer information of deep neural network Res Net-50 and analyzing the advantages and shortcomings of traditional optimization algorithm Adam.The CSP cross-stage ratio is determined by experiments,and the ablation experiments on the experimental wood surface defect data show that the classification accuracy of the improved Res Net-50 model is 88.55%,which is better than those of the traditional Res Net-50(86.77%)and other traditional classification networks.(3)To address the problems that the accuracy of the traditional target detection network Faster RCNN is not good enough and the overlapping defects of the same kind are easily missed,this paper improves the feature extraction of backbone network,loss function and non-maximal suppression of the traditional Faster RCNN.The improved Res Net-50 in(2)is used as the backbone network,and Focal Loss is adopted as the loss function,and Soft NMS is used as the non-maximal suppression.Experimental results indicate that the improved Faster RCNN has a great improvement on the experimental wood surface defect dataset.