Research On Methods For Robust Texture Feature Extraction Based On Local Binary Pattern

Texture is widely present in the real word and it is an important visual clue for humans to perceive different objects.The extraction of effective texture features is one of the fundamental tasks in image processing,computer vision and pattern recognition.As a classical local feature extraction method,Local Binary Pattern(LBP)is widely applied in texture classification,face recognition and other fields due to its simple theory,low computational complexity and good gray-scale invariance.However,the existing LBP methods have poor robustness to noise,illumination and scale change and can not obtain satisfactory classification performance.Therefore,the goal of this thesis is to improve the discriminative power and enhance the robustness of LBP features.The main contents are as follows:1.The existing LBP-based methods can not capture the relationship between neighboring pixels and lack of feature description between non-local pixels.In view of these problems,this thesis proposes a feature extraction method based on local and nonlocal patterns.Firstly,this method encodes the neighboring pixel gray order relationship by a Local Grouped Order Pattern(LGOP)based on the dominant direction.Then,the gray difference between the neighboring pixels and several anchor points(based on global image statistics)is encoded by a Non-Local Binary Pattern(NLBP).Finally,the joint center pixel coding is used to form the texture feature representation.The experimental results show that the proposed method can effectively improve the texture classification performance under various imaging conditions.2.The features extracted by existing LBP-based methods have weak discriminative power and poor robustness.To improve LBP,this thesis proposes three local binary coding texture feature extraction methods based on image decomposition.(1)Multi-Scale Sectored Local Binary Pattern(MSLBP)obtains low-frequency,positive high-frequency and negative high-frequency images at multiple decomposition levels through pyramid decomposition and thresholding processing techniques.Then,the feature codes of the decomposed images are calculated by Sectored LBP(SLBP)based on local average operation.Finally,the texture feature is obtained by joint coding across frequency bands and by histogram weighting across decomposition levels.(2)Enhanced Multi-Scale Sectored Local Binary Pattern(EMSLBP)calculates multi-scale decomposition images through pyramid decomposition and polarity separation operations,and performs SLBP coding on the decomposition images.The multi-scale texture feature representation is constructed by joint encoding of central pixels across frequency bands.(3)Local Binary Pattern Coding Across Scales,Frequency Bands and Image Domains(CSFD-LBP)obtains multi-scale low-frequency and high-frequency images by Gaussian filtering and image subtraction.Meanwhile,multi-scale gradient images are computed based on Gaussian derivative filtering.Then,the LBP codes of the low frequency images,high frequency images,and gradient images are calculated.Finally,the feature histogram representation is constructed by joint LBP coding across scales,frequency bands and image domains.The experimental results show that these three methods studied achieve good classification performance under noise-free and Gaussian noise conditions.