HJDLBP: A novel texture descriptor and its application in face recognition

Local binary pattern (LBP) is a simple and computationally efficient texture descriptor which has been attracting many attentions since its introduction; due to the extensive research done in this regard, diverse variants of LBP have been introduced in recent years. While original form of this operator encodes structures like spots, edges, and corners in form of a binary code, a more recent type of LBP called high order directional derivative LBP (DLBP) reveals some alternative structures such as convexities and concavities. Even though these structures are important features in the images, another significant consideration is the relationship between them. For instance, there is a high probability that an edge structure be present near another one. In this paper, we have introduced a novel texture descriptor named HJDLBP (high order joint DLBP) which is able to encode relationships between micro patterns in addition to the prevalent structures. To evaluate the proposed descriptor, we have considered two renowned JAFFE and YALE facial image databases and then exploited the proposed texture descriptor for face recognition issues. The experiments are implemented in software in the following manner: as a first step, the face part of each image is segmented from its background using Viola and Jones algorithm. Afterward, the micro patterns and relationships between them are extracted from rectangularly partitioned face images; and their histograms are constructed as well. Finally, a group of SVMs are trained for classification. We have compared obtained results using the new operator with the results attained when conventional LBP and high order DLBP are applied for feature extraction from image blocks. The comparative results show the efficacy of the proposed operator as a texture descriptor.