One Missed Relief Feature Selection of Extended Local Binary Pattern for Texture Classification

Image processing plays a major role in a wide range of artificial intelligence applications. Among the various types of images, texture images present greater complexity due to their intricate patterns and variability. Numerous techniques have been developed for feature extraction from texture images, with the Local Binary Pattern (LBP) being one of the most widely used. A notable extension of this method is the Completed Local Binary Pattern (CLBP), which enhances classification performance by combining three distinct feature histograms. However, this approach significantly increases the feature dimensionality, potentially leading to computational inefficiency. In this paper an Extended Completed Local Binary Pattern (ECLBP) method is proposed that provides more designative features than CLBP. ECLBP provides higher number of discriminative features than most of the other descriptors such as CLBP. Therefore, to decrease this number, this study introduces two novel feature selection algorithms based on an improved variant of the Relief feature selection technique. Relief is a feature selection technique based on weighting discriminative features for binary class problems. ReliefF has been proposed for multi-class data. It uses the near-hit data of the same class and all near-miss data of other classes. In this research, two novel versions of ReliefF are proposed that only use one-miss data of other classes. The proposed method is named One Missed Relief (OM_Relief) because it uses only one nearest data sample of all missed data of other classes. The proposed approaches are filter-based and supervised feature selection, employing a score-based feature selection method that incorporates weighted Relief. The implementation of the proposed method on eight various datasets, including general textures, coral reef, face, and virus textures datasets, indicates the ability of the proposed method to achieve a higher classification rate compared to some state-of-the-art methods.