Local Ternary Patterns and Maximum Bipartite Matching for Face Recognition

Local Binary Pattern (LBP) has been the successful feature descriptor used for face recognition. The basic idea in this method is to convert from an intensity space to an order space where the order of neighboring pixels is used to create a monotonic change illumination-invariant code for each point in the image. A drawback for this method, however, is that, in homogenous regions, the order of the pixel with respect to its neighbors is quite noisy. In this paper, we propose to use a third value which indicates if two pixels are similar in value. We also propose to match these patterns using maximum bipartite matching rather than histogram matching. Significant performance boost was found when compared to LBP and other standard methods like Principal Component Analysis (PCA) and Linear Discriminant Analysis (LDA). We present our results on Extended Yale and FERET datasets.