Study On Texture Representation Based On Local Binary Patterns

The texture of a surface directly reflects properties of the object. The texture in animage supplies a wealth of information for expressing the object. The description of thetexture has been one of the focuses and difficulties in the computer vision field all the time.In recent years, the local binary pattern as one of texture features has been widely appliedto computer vision in various fields. This paper has deeply studied the local binary patternsfor texture classification.Most descriptors are vulnerable to the illumination changes. Analyzing the effect on thelocal binary pattern methods and focusing on the reflecting illumination change, this paperproposes the gradient local binary patterns. The method employs the gradient informationto change the comparison of a local neighborhood and construct local binary patterns. Theproposed method is useful for the images with linear transformations.In addition, the rotation changes of textures need the local descriptors to be rotationinvariant. At present, some local binary pattern algorithms have good solutions to the rota-tional images. Studying the design ideas of these algorithms, three strategies to build localbinary patterns have been summarized. Therefore, three local binary pattern algorithms havebeen proposed according to the summarized strategies. Multi-ring local binary patterns isone of the proposed approaches. The multi-ring local binary patterns employ the combiningstrategy to merge the information of sampling points located at the same ring, and buildthe rotation invariant local descriptor. Local rank binary patterns follow a strategy of thetransformation to sort the sampling points, and use the position of central point in the sortedsequence to represent the local structure. Altering the first sampling point changes the posi-tion of the first sampling point of the conventional local binary patterns to build features indifferent orientations. The method of the altering the first sampling point could make mostlocal binary patterns achieve the rotation invariant texture classification with an exhaustivelysearching strategy.In order to improve the discriminative power of the local binary patterns for texture classification, this paper proposes the multi-structure local binary patterns. The proposedmethod combines the isotropic sampling and the anisotropic sampling to expand the localbinary patterns, uses the image pyramid to extract micro structures and macro structures oftextures. Appropriate weights have been assigned to different extracted features accordingto their contributions on texture description. Moreover, the proposed approach has goodexpansibility, and can be applied to some basic local binary pattern algorithms to improvetheir discriminatory power.The dictionary learning and local binary pattern are two popular methods of textureclassification. The two methods have many similarities in describing objects, feature extrac-tion and so on. In this paper, the authors have analyzed the two types of algorithms, andfound that the local binary pattern algorithm was a special dictionary learning algorithm. Inorder to remove the particularity of local binary patterns, two ways of building dictionariesare used to improve the local binary pattern algorithm. The local binary patterns based onpattern dictionary learn the effective patterns from rotation invariant local binary patterns.The local binary patterns based on feature dictionary employ the histograms of local binarypattern in small regions to build the effective dictionary. The local binary patterns based onfeature dictionary combine the advantages of dictionary learning and local binary patterns,and show good abilities on texture description.At last, the presented work is summarized. According to the imperfect aspects, futurework has also been analyzed and discussed.