Soft local binary patterns

Abstract

Local Binary Pattern (LBP) is known as one of the most effective local descriptors for image recognition. It is invariant to monotonic gray-scale changes of the image. Local neighborhood information is gathered for each pixel of the image, and a binary code is generated by comparing its value with the value of the center pixel. Then a histogram of binary code is created by counting up the occurrences of the different binary patterns. In this paper we propose an extension of the original LBP by using a soft thresholding function instead of the hard thresholding function using in the original LBP. Then we construct the histogram by voting the weights calculated depending on the distance between the extracted feature vector and the binary vectors. By using the proposed Soft LBP, we can extract information on the differences between the value of the center pixel and the value of the neighboring pixels. This means that the details of the textures can be included in the extracted features. To confirm the effectiveness of the proposed Soft LBP, we have performed the experiments on face recognition and face expression recognition. The results shows that the proposed Soft LBP gives better recognition rates than the original LBP and and the co-occurrence of adjacent local binary pattern and is comparable with the Soft Histogram LBP.