A method for heterogeneous face image synthesis

A novel learning based framework for efficient heterogeneous faces synthesis is proposed. Based on the same spectral distribution of each modality, a statistical probability model is developed for the mapping learning problem between two groups of facial appearances, instead of the traditional linear regression model. Furthermore, in order to eliminate the influences of facial structure and spectrum on the training model, a 3D model is applied for facial pose rectification and pixel-level alignment, and Difference of Gaussian(DOG) filter is adopted to normalize the image intensities. Experiments on HFB database demonstrate that this scheme provides promising results both in image representation and in face recognition.