Local Descriptor Methods For Textyre Classification And Leaves Recognition

Texture classification is a basic issue of image processing and computer vision, and has an extensive application background in science research and engineering technology, e.g., remote sensing, biomedical image analysis, biometrics, image retrieval. Extracting the texture feature is the foundation of texture classification. In the past decades, the researchers have got a consensus about the conception of texture: Macroscopic texture can be regard as the repeats for a large number of local microcosmic textural patterns, and it can be characterized by the statistic of the local texture patterns. Recently, local texture descriptor has become a popular research topic of texture classification. Some efficient local texture descriptors have been proposed, e.g., Local Binary Pattern (LBP). Generally, texture images captured in the real world may have obvious illumination, orientation, and scale variations. How to enhance the discriminative capability of local descriptors and keep them invariant to illumination and viewpoint is still worth further research.In this paper we proposed several new local texture descriptors. We try to find the key discriminative information of local texture. In conventional LBP theories, the local texture information is divided into two orthotropic components:the local gray-value difference distribution and the local micro-structure. The micro-structure is thought to be the main information of local texture. Firstly, we proposed a new local operator which totally discards the structural information of the LBP, and the experimental results demonstrate that the key discriminative information of local texture is the distribution of local gray-value difference. Then we try to extract more information about the local gray-value difference by increasing the local quantization levels. Secondly, we presented a robust framework of the LBP, and the proposed RLBP can be more insensitive to noise and illuminations. Thirdly, we proposed another interesting local descriptor, named local connection code, to extract the rotation invariant micro-structure information. Lastly, we presented two effective leaves recognition methods by using two local texture operators.