The local binary pattern approach to texture analysis - extensions and applications

This thesis presents extensions to the local binary pattern (LBP) texture analysis operator. The operator is defined as a gray-scale invariant texture measure, derived from a general definition of texture in a local neighborhood. It is made invariant against the rotation of the image domain, and supplemented with a rotation invariant measure of local contrast. The LBP is proposed as a unifying texture model that describes the formation of a texture with micro-textons and their statistical placement rules. The basic LBP is extended to facilitate the analysis of textures with multiple scales by combining neighborhoods with different sizes. The possible instability in sparse sampling is addressed with Gaussian low-pass filtering, which seems to be somewhat helpful. Cellular automata are used as texture features, presumably for the first time ever. With a straightforward inversion algorithm, arbitrarily large binary neighborhoods are encoded with an eight-bit cellular automaton rule, resulting in a very compact multi-scale texture descriptor. The performance of the new operator is shown in an experiment involving textures with multiple spatial scales. An opponent-color version of the LBP is introduced and applied to color textures. Good results are obtained in static illumination conditions. An empirical study with different color and texture measures however shows that color and texture should be treated separately. A number of different applications of the LBP operator are presented, emphasizing real-time issues. A very fast software implementation of the operator is introduced, and different ways of speeding up classification are evaluated. The operator is successfully applied to industrial visual inspection applications and to image retrieval.