Pyramid and multi kernel based local binary pattern for texture recognition

Abstract

In this article, a novel pyramid and multi kernel based method is proposed to increased success of the local binary pattern (LBP). Signum, ternary and quaternary binary feature extraction functions are used together and these are utilized as mathematical kernel of the LBP. In order to extract features in depth, pyramid model is used. Texture images are resized in the 4 levels to create pyramid. Finally, 5120 features are extracted from each level. In the feature reduction phase, principle component analysis is considered and linear discriminant analysis is utilized as classifier. To obtain numerical results, UIUC, Outex and USPTex datasets were used. The proposed method was compared to the other state of art texture classification methods. The recognition rates were calculated as 96.10%, 89.90% and 97.30% for UIUC, Outex and USPTex respectively. The robustness tests were performed using the Gaussian and salt and pepper noises. The best accuracy rates of the noisy images were calculated as 79.5% and 94.3% respectively. The experimental results proved the success of the proposed method.