Linear Discriminant Analysis(LDA) is the well-known method in pattern recognition,and researchers usually use PCA+LDA instead of LDA due to the Small Sample Size Problem(SSSP) in image recognition.A Multi-band Linear Discriminant Analysis(MBLDA) is proposed,by which LDA is established on the whole samples space and SSSP is resolved.Based on MBLDA,the data loss using PCA is avoided,the dimension of discrimination feature extracted is very low and the recognition performance is advanced.Its recognition rate exceeds PCA,LDA,or PCA+LDA largely.Some experiments on ORL and NUST603 face database demonstrate that the proposed method is effective.