Learning cognitive manifolds of faces
Liyuan Li, Bappaditya Mandal, Cheston Tan, Joo‐Hwee Lim · 2017
Inspired by the studies in psychology and neuroscience, we propose a computational model of cognitive face representation that mimics the mechanism of human face perception. We propose to learn two separate manifolds for facial identity and facial expression. A new t-SNE algorithm is designed to find a way to achieve the learning purpose and a spectral regression algorithm is implemented accordingly to learn the cognitive manifolds of neutral and smiling faces from real face images. The association between the two manifolds is established by a weighted k-NN fusion algorithm. An iterative linear algorithm is derived to refine the association between the two manifolds. We evaluate our model on two large datasets containing real-world images of neutral and wide smiling faces. The experimental results show both the accuracy of the cognitive manifold representation and the superiority over Eigenfaces and Fisherfaces on the tough task of matching a wide smiling face to a neutral face of an individual.