The invention discloses a face recognition method comprising the following steps: extracting P face images from a face set, dividing the remaining images in the face set into n categories of face images and taking the remaining images as a training sample library, and pre-classifying the P face images based on category mean information in the training sample library; after classification, performing low-rank recovery on each of the P face images and an image of corresponding category to get low-rank recovery images and sparse error images; using the training sample library to train an extreme learning machine based on the principle that the low-rank performance of the low-rank images is robust to noise; and finally, by taking the low-rank recovery images corresponding to the P face images as a test sample library and the extreme learning machine as a standard, building the mapping relationship between the P face images and face labels obtained after training, and performing face recognition on the low-rank recovery images corresponding to the P face images. The recognition robustness of the extreme learning machine is improved, and the computational complexity is reduced.