Subspace modeling plays an important role in face recognition. Independent Component Analysis (ICA), a multivariable statistical analysis technique, can be seen as an extension of traditional Principal Component Analysis (PCA) technique, which addresses high order statistics as well as second order statistics. In this paper, a new scheme of subspace-based representation called Discriminant Independent Component Analysis (DICA) is proposed, which combines the strength of unsupervised learning of ICA and supervised learning of Linear Discriminant Analysis (LDA), and efficiently enhances the generalization ability of ICA-based representation method. Based on DICA subspace analysis, a set of optimal vectors called "discriminant independent faces" are learned from face samples. The effectiveness of our method is demonstrated by performance comparisons with some popular methods such as ICA, PCA, and PCA+LDA. On the large scale database of ⅡS, significant improvements are observed when there are fewer training samples per person available.