The invention provides a discriminant sparse preserving embedding method for unconstrained face recognition, 1) calculating a sample reconstruction relation matrix W, when calculating the sparse reconstructed relation of samples, the class label is introduced to construct the intra-class reconstructed relation matrix and inter-class reconstructed relation matrix respectively, and the intra-class and inter-class compactness constraint is added in the sparse reconstructed stage, which effectively increases the reconstructed relation between the samples to be tested and the same kind of samples,and weakens the reconstructed relation between the samples to be tested and the heterogeneous samples. 2) When calculating the low-dimensional projection matrix P and the low-dimensional projection matrix, the global constraint factor is added, which not only considers the local sparse relation of the sample, but also considers the global distribution characteristic, further weakens the disturbance of the heterogeneous pseudo-nearest neighbor sample to the low-dimensional projection, and more accurately excavates the essential structure of the low-dimensional manifold hidden in the complex redundant data; 3) The low-dimensional linear mapping of high-dimensional sample data is realized, which greatly improves the accuracy of face recognition in unconstrained environment.