In this paper, a recently developed local feature descriptor, namely directional local binary patterns (DLBP), was first proposed for palmprint recognition. Compared with local binary patterns (LBP) and directional binary code (DBC), DLBP contains more information on both edge and texture. A cascade structure using AdaBoost algorithm is then used to reduce the feature dimension of DLBP and computational costs of classification. The proposed approach was applied to fuse multispectral palmprint images captured under red, green, blue and near-infrared (NIR) lighting sources for personal identification. Experimental results suggest that the proposed algorithm performs much better than DBC, LBP and PalmCode in identifying palmprint images captured using different illuminations. When fusing the multispectral images, the proposed approach has also been shown to achieve higher accuracy than other methods in literature such as QPCA (Quaternion PCA) and QDWT (Quaternion Discrete Wavelet Transform).