A Local Directional Pattern Variance (LDPv) Based Face Descriptor for Human Facial Expression Recognition

Automatic facial expression recognition is a challenging problem in computer vision, and has gained significant importance in applications of human-computer interaction. This paper presents a new appearance-based feature descriptor, the Local Directional Pattern Variance (LDPv), to represent facial components for human expression recognition. In contrast with LDP, the proposed LDPv introduces the local variance of directional responses to encode the contrast information within the descriptor. Here, the LDPv representation characterizes both spatial structure and contrast information of each micro-patterns. Template matching and Support Vector Machine (SVM) classifier are used to classify the LDPv feature vector of different prototypic expression images. Experimental results using the Cohn-Kanade database show that the LDPv descriptor yields an improved recognition rate, as compared to existing appearance-based feature descriptors, such as the Gaborwavelet and Local Binary Pattern (LBP).