Weber vector local pattern

Feature extraction of face images, especially those with significant illumination, occlusion and facial expression changes, is a hot topic in the field of computer vision. This paper presents a novel local image descriptor called Weber vector local pattern (WVLP). WVLP uses the salient variations within an image to simulate the pattern perception of human beings. Different from most WLD-like methods which describe a pixel with the isolated salient variations and direction components, WVLP introduces the Weber vector to deduce two compound features, i.e., the salient variations and the direction feature, which work together to make the feature images more discriminative. During the first feature transformation, two sizes of neighborhoods are used to capture the structural features, and four feature images were output. After that, an LBP-like trick, named neighbors to center difference binary pattern (NCDBP) is used to further extract the high-level complex structure information in a large range of neighborhood. Experimental results on the several face recognition databases show that WVLP outperforms other compared descriptors, even in the database with challenges of illumination, facial expression, and occlusion changes.