Research On Surface Defect Detection Of Wood Based On Digital Image Processing

Wood defects will reduce the external quality and internal strength of wood,and affecting the processing of wood production process.At present,the method of artificial judgment is usually used to detect wood defects.The processing of the wood is a repetitive process,which is prone to visual fatigue after prolonged detection for a long period of time,resulting in a decrease in the quality of the test.At the same time,for some processing technology,such as parts cutting,artificial methods difficult to quickly and accurately count the distribution and category of wood defects.In order to realize the requirement of automatic wood processing,the purpose of this paper is to locate and classify the surface defects of wood based on digital image processing technology.The specific research contents are as follows:(1)Preprocessing is used to improve the contrast of wood images.The Ostu algorithm was used to segment the wood surface defects of three common categories.Through the mathematical morphology,a more smooth and complete defect area was obtained.For the background complex examples,three channels based on the HSI color model are used as the characteristics to segment the wood surface defects through K-mean.For the multi-defect samples,K-mean achieves better segmentation effect compared with the Ostu algorithm,but there are still problems such as incomplete segmentation and misclassification.(2)Based on the LBP training Adaboost cascade classifier,the wood surface defects are located and compared with the Adaboost cascade classifier trained by Haar and HOG characteristics.The experimental results show that the cascade classifier trained based on LBP has achieved better detection results.The recall rate is 96.9% and detection accuracy rate is 96.2%.At the same time,the cascade classifier trained by LBP is more accurate forwood surface defects,with the shortest test time,less memory for testing,and better robust performance for complex scenes.(3)Extract the gray covariance matrix vector of the sample to train the classifier.When KNN is used as a classifier,the minimum recognition rate is the lowest when the nearest neighbor number is 3.When using SVM as a classifier,the polynomial is used as the kernel function with the lowest recognition rate.When using BP neural network as a classifier,the deeper hidden layer number and the number of neurons can improve the classification performance of the model,and the minimum error recognition rate is 12.5%.(4)According to the problem that the characteristics of traditional classification methods are difficult to be selected,the surface defects of wood are classified by convolution neural network.Increasing the number of convolutions per layer,using a smaller convolution kernel,using a deeper network layer,the use of Relu function as an activation function could improve the classification performance of the model.At the same time,adding the Batch Normalization layer to the model could make the training of the model easier to achieve a steady state.Through the data amplification technology,can reduce the training data generated by the over-fitting problem.Through the continuous adjustment of the model and the use of data augmentation,etc.,the model of the smallest misrecognition rate of 2.3%.(5)Through the combination of Adaboost cascade classifier and CNN algorithm,the localization and classification of wood defects are realized.The experimental results show that the recall rate of the algorithm is 94%,the correct rate of detection is 99%,and the correct rate of classification is 97.9%.At the same time,the average test time of the algorithm is 102 ms,the test memory is 120 M,which meets the requirement of real-time online detection.