The invention discloses a single-sample face recognition method based on a self-adaptive virtual sample generation criterion. The single-sample face recognition method mainly solves the problem of low face recognition rate in the prior art, and is implemented by the steps of: 1, selecting face images and dividing the face images into training sample sets and testing sample sets; 2, carrying out singular value decomposition on a training sample, and reconstructing a new training sample image according to the decomposed basic images; 3, constructing a virtual training sample image by combining the training sample image with the new reconstructed image, and partitioning the training sample image and the virtual training sample image to form a block training sample set; 4, training an optimal projection space by utilizing block training samples; 5, partitioning a testing sample in the same manner, and projecting the partitioned testing sample to the optimal projection space to obtain block sample features; 6, and partitioning block testing samples according to the block sample features, and acquiring a final recognition result by means of a maximum voting rule. The single-sample face recognition method reduces the loss of authentication information in the face recognition, improves the face recognition rate, and can be applied to recognition of ID cards, driver's licenses and passports.