A Research Of Woodboard Recognition Based On Visual Texture And Machine Learning

With the development of image processing and artificial intelligence,intelligent manufacturing has become more and more popular.Nowadays,people's quality of life has been greatly improved,and the requirements for home comfort have become higher and higher,especially for the appearance of wooden furniture.To satisfy the requirements of furniture customization,the manufacturer needs to classify the wooden board based on the surface characteristics and use the same kind of wooden board to make the furniture according to the customers' orders.In this situation,traditional artificial means cannot adapt to the rapid development of the smart home industry.To overcome such shortcoming,this thesis aims at propose three recognition methods adaptive to different computing powers.On the one hand,in case of low computing power,this paper,adopting the Gray-Level Co-occurrence Matrix(GLCM)based texture feature extraction method,employs Mahalanobis Distance(MD)and the Support Vector Machine(SVM),respectively,for automatic wood recognition.On the other hand,in case of high computing power,a Convolutional Neural Network(CNN)with high recognition accuracy is designed and tested by read data.The main research work of this thesis includes:a)The acquisition of and pre-processing of wood board image data.The wood panel images utilized in this work were collected in the factory by a scanner.In order to improve the quality of the board image,the pre-processing,including smoothing and de-noising,are firstly performed on the image.The methods of parameter setting and feature extraction are then discussed.After a detailed analysis of existing methods for feature extraction,we select the GLCM based technique as our feature extraction method,since GLCM is robust against position drift,brightness change,color variation,and so on.In this work,we extracted four kind of features based on GLCM to describe and measure the texture of the wooden board.b)Recognition methods suitable for low computing power.The basic principles of MD and SVM are elaborated,laying foundations for the future experiments.Numerical results suggest than SVM outperforms MD based on the four kinds of features extracted from GLCM.c)Recognition methods suitable for high computing power.To further improve the accuracy obtained by MD and SVM,this work further applies CNN to wooden board recognition.Based on the existing theory and method of CNN,we adopted the framework of TensorFlow,designed and constructed a CNN consisting of eleven hidden layers.After training and fine-tuning of the parameters,the proposed CNN can achieve an accuracy as high as 98.8%,which satisfies the requirements of the application.The three recognition methods,namely,MD and SVM for low computing power scenarios,and CNN for high computing power scenarios,are not only of theoretical value,but of important applicable importance for furniture manufacturers aiming at reducing cost and improving accuracy.