The well-known eigenface method uses a single eigenspace to recognize faces. However, it is not enough to represent face images with large variations, such as illumination and pose variations. To overcome this disadvantage, many researchers have introduced multiple eigenspaces into face recognition field. But most of these methods require that both the number of eignspaces and dimensionality of the PCA subspaces are a priori given. In this paper, a novel self-organizing method to build multiple, low-dinensinal eigenspaces from a set of training images is proposed. By eigenspace-growing in terms of low-dimensional eigenspaces, it completes clustering images systematically and robustly. Then each cluster is used to construct an eigenspace. After all these eigenspaces have been grown, a selection procedure eigenspace-selection is used to select the ultimate resulting set of eigenspaces as an effective representation of the training images. Then based on these eigenspaces, a framework combined with neural network is used to complete face recognition under variable poses and the experimental result shows that our framework can complete face recognition with high performance.