Statistical Feature Selection for Image Texture Analysis

Texture is one of the visual features used in Content Based Image Retrieval (CBIR) to represent the contents of the image with respect to the characteristics brightness, color, shape, size, etc. Texture is a property that represents spatial distribution of an Image. Texture can be defined as a repetition of an element or pattern in a problem space. Texture analysis can be used for classification, segmentation, feature extraction and shapes determination [1]. Several methods are proposed to describe the texture of an image . The four major categories of these are, statistical, geometrical, modelbased and signal processing [ 2]. Statistical approaches represent the texture by the statistical distribution of image grey values and compute statistics from spatial distribution of the local patterns [2]. These are classified into first-order, second-order and higher-order based on the number of pixels used for a local pattern. Individual pixels are considered in first-order statistics to compute features like average, variance. In second-order statistics, frequency of each pair of pixels is used to extract the features of the image. second-order statistical methods have provided more prominent results when compared to other methods like transform-based and structural methods [3]. Advances in medical imaging technologies such as Computed tomography (CT), Magnetic resonance (MR), Ultrasound imaging etc. have made the use of computers compulsory in diagnosis and treatment planning. Image feature extraction is vital in Content Based Image Retrieval (CBIR) for analysis and interpretation of images [4]. It helps radiologists in the early detection of abnormalities and treatment planning. The quality of diagnosis often depends on how accurately the various structures in the image are identified. These can be identified with various texture methods. Various statistical approaches for texture representation are Co-occurrence matrices, fractal model, Tamura feature, world decomposition and so on . The co-occurrence matrices are second order statistics to measure texture properties of an input image. Various cooccurrence matrices in the literature are Graylevel cooccurrence Matrix (GLCM) [5], Graylevel Aura Matrix (GLAM)[6], Graylevel Run Length Matrix (GLRLM)[7], Graylevel Neighbor Matrix (GLNM) [8]and so on. This paper presents details about various second-order statistical matrices used to extract features of the input image.