In this paper, a novel scheme is proposed for face recognition or authentication against illumination, and expression variation using multimodal face features. First, a sub -image in low-frequency sub-band is extracted by a wavelet-based transform to preserve the invariant data and reduce the dimensionality. The reduced sub-image LL is partitioned into four parts for representing local features and reducing illumination effect. Sub-image LL is reduced again to obtain a full face in a smaller scale for globally representing the whole face image. Five modal feature spaces are constructed. The most discriminant common vectors (DCVs) for each classifier are found. A weighted summation is performed to fuse the five classified results. Experimental results were conducted to show that the proposed scheme is superior to other methods in terms of recognition rate and authentication rate.