Noise-Robust Local Ternary Pattern Center for Noisy Texture Classification

Local Binary Pattern (LBP) has been widely used for texture analysis in machine vision and image processing. However, most of these descriptors provide a high number of features that are sensitive to noise and rotation. In this paper, a version of LTP is proposed that is named Local Ternary Pattern Center (LTPC). This descriptor provides noise-robust and low feature numbers with a better classification rate than LTP and advanced LBP versions for noisy textures. The proposed method has three steps. First a criterion is proposed that estimates the quality of the texture. According to the estimated quality of the texture, an average kernel is applied to the noisy texture. The more level of noise the more number of applying average kernel must be convolved to each texture. Convolution is a time-consuming operation, so here, a fast average method is proposed that can do this operation around 2 to 3 times faster than the traditional convolution method. In the second step LTPC is used to extract the features of filtered textures. The third innovation of this paper is related to the mapping of the features. Most of the advanced LBP versions used the riu2 mapping method to decrease the number of features and make them rotation-invariant. In this research, a novel mapping method is proposed. It has all advantages of the riu2 and it provides features that they are more robust to noise than riu2. The implementations on different datasets such as Outex suits (TC10, TC12(t), TC12(h), TC13), CUReT, UIUC, USPTex, Virus, RSMAS, KTHTIPS2, UMD and ORL indicate that the proposed method can provide very high discriminative and noise-robust features. The higher the level of noise the better classification accuracy can be obtained by the proposed method compared to some state-of-the-art LBP versions. Furthermore the classification rate of proposed method is better than some deep neural network results. Whereas, in term of speed the local descriptors provide very shorter operation time than most of the deep learning methods.