Adaptation of Local Binary Patterns toward Robustness to Gaussian Noise

Local binary patterns represent very powerful feature for image classification. It evolved from the original model to many modifications and adaptations. However, almost all of derived models are based on the basic idea to code each pixel's 3×3 neighborhood binary. In this paper we propose modification of this idea in order to increase its robustness to Gaussian noise. Instead of calculating differences on 3×3 neighborhood regarding central pixel, we apply calculation of differences regarding average pixel intensity on that neighborhood. This kind of averaging improves classification performances with respect to noise. Proposed method is further tested for classification on two publicly available texture datasets and obtained results, with and without noise addition, prove our assumptions.