Texture Image Recognition Based On Two-Dimensional Local Binary Pattern

Image feature extraction plays a non-negligible role in image processing,computer vision and pattern recognition.Effective image features provide more essential image information.However,unfavorable factors such as illumination,translation,rotation,and stretching can cause geometric of the image during imaging.As an effective image feature extraction operator,Local Binary Pattern(LBP)can resist the distortion caused by problems such as illumination,grayscale and rotation,and stably extract image feature.LBP is used in image recognition,texture classification,image analysis,image compression,image fusion and face recognition,especially in the field of texture image recognition.In recent years,research scholars mainly discuss and study the one-dimensional characterization operator of LBP,this thesis proposes a two-dimensional local feature description operator to enrich the context information of features;On the other hand,after obtaining the feature description operator of the texture image,this thesis proposes the twostages classifier image recognition method,the specific research is as follows:1.At present,the research of local binary pattern mostly changes the coding pattern methods,pattern selection method,domain topology,et al.These methods only consider the information of the single pixel.In this thesis,the local binary feature extraction operator based on two-dimensional binary pattern is proposed.For the original LBP operator without considering the correlation pixels,a two-dimensional feature extraction operator is designed to enrich the correlation and context information between pixels.In the process of constructor construction,this thesis adopts the rotation invariant uniform pattern of the original LBP for feature description,and introduces the sliding window concept to obtain more context information by changing the size of the sliding window.Two-dimensional histogram is used to statistical feature descriptors,the concept of weights is introduced,and dominant features that are more favorable for texture image recognition are selected.This thesis uses the Local Threevalued Pattern(Local Three-valued Pattern,LTP)to verify the effectiveness of the twodimensional idea.2.Aiming at how to classify after feature extraction,a classification idea of two-stages classifier is proposed.Among many classifiers,there are linear classifiers and nonlinear classifiers.The most commonly used classifiers are nearest neighbor classification(k-Nearest Neighbor,KNN),decision tree classification,naive Bayes classifier,and Support Vector Machine(Support Vector Machine,SVM).The application of the above classifiers is limited.In this thesis,a novel two-stages classifier are proposed for the image classifiers.The secondlevel classifier improves the final images Correct Classification Percentages(Correct Classification Percentages,CCPs)by correcting the prediction values of the first classifier.