This paper presents a subspace algorithm called block independent component analysis (B-ICA) for face recognition. Unlike the traditional ICA, in which the whole face image is stretched into a vector before calculating the independent components (ICs), B-ICA partitions the facial images into blocks and takes the block as the training vector. Since the dimensionality of the training vector in B-ICA is much smaller than that in traditional ICA, it can reduce the face recognition error caused by the dilemma in ICA, i.e. the number of available training samples is greatly less than that of the dimension of training vector. Experiments on the well-known Yale and AR databases validate that the B-ICA can achieve higher recognition accuracy than ICA and enhanced ICA (EICA).