Novel Fuzzy Kernels Based Local Binary Pattern And Local Graph Structure Methods

Abstract

Local descriptors are the most effective textural image recognition methods. Local descriptors generally consist of two phases. These are binary feature coding and histogram extraction phases, and they often use the signum function for the binary feature extraction. In this article, new fuzzy-based mathematical kernels are proposed for binary feature encoding in local descriptors. Fuzzy kernels consist of membership degree calculation and coding these membership degrees. In order to calculate membership degrees, four fuzzy sets are utilized. The proposed fuzzy kernels are considered as binary feature-extraction functions, and a novel textural image recognition architecture is created using these fuzzy kernels. These architecture phases are; (1) binary feature coding with fuzzy kernels, (2) calculating lower and upper images, (3) histogram extraction, (4) feature reduction with maximum pooling, (5) classification. In the classification phase, a quadratic kernel-based support vector machine (SVM) classifier is utilized. The presented fuzzy kernels are implemented on the Local Binary Pattern (LBP) and Local Graph Structure (LGS). 16 novel methods are presented using fuzzy kernels for each descriptor. In this article, LBP and LGS are used, and 32 novel fuzzy-based methods are proposed to improve recognition capability. 3 facial images and 3 textural image datasets are used to evaluate the methods' performance. The experimental results clearly illustrate that the fuzzy kernels based LBP and LGS methods have high facial and textural image recognition capability.