Translated Abstract
Texture classification is a basic issue of image processing and pattern recognition, and widely used in engineering technology, e.g., remote sensing image analysis, medical image analysis, biometrics, industrial detection and so on. The extraction of texture features is the basis of texture classification. Local texture descriptor has become a popular feature extraction method of texture classification. In recent years, many local texture descriptors have been proposed. Among local descriptors, LBP has emerged as one of the most prominent descriptor and has attracted increasing attentions in the field of texture classification due to its outstanding advantages. In general, texture captured in the real environment may have rotation, illumination, and viewpoint variations. How to improve the discriminative capability of LBP and keep LBP invariant to rotation, illumination and viewpoint changes is still needed further research. Therefore, this paper focuses on improving the performance of LBP in texture classification and the main contents of this paper are as follows.
(1) In consideration of the quantization process in LBP is coarse, we introduce the vector quantization method into quantization process and propose Local Vector Quantization Pattern (LVQP). In a local region, LVQP takes the difference vector between the neighbor pixels and the central pixel as a whole, and directly uses the difference vector for quantization. To achieve rotation invariance, we proposed a circulation search method in the quantization process. Experimental results show that LVQP is robust for noise and illumination changes and have better classification accuracy than the related state-of-the-art methods.
(2) A novel feature-based local binary pattern (FbLBP) is proposed. In FbLBP, the difference vector between the neighbors and the central pixel is divided into two vectors: the sign vector and the magnitude vector. The uniform LBP is used to encode the sign vector and the features of the magnitude vector is used to represent the magnitude vector. The features of the mean and the variance provide complementary information to LBP. Furthermore, the multi-scale FbLBP is proposed to extract features from different sampling radiusesto fully describe the local texton structure. Experimental results show that the proposed FbLBP significantly outperformed other related methods as well as kept a smaller feature dimension.
(3) A novel local ring-likebinary pattern named LRBP is proposed. The algorithm converts the rectangular neighborhood into multi-ring neighborhood. LRBP extracts features from multi-ring neighborhood. A novel feature extraction method that describes the feature of whole sampling neighbors is proposed. Furthermore, the uniform pattern is not always occupy the majority in the complex texture images. An extended definition of the uniform pattern is proposed to make full use of the binary patterns. LRBP which uses the extended uniform patterns is called Extended LRBP (ELRBP). Experimental results show that LRBP and ELRBP have stronger discriminative ability than other related methods.
Corresponding authors email
