The invention discloses a face recognition method based on aggregate loss deep metric learning. The method comprises the steps of 1), preprocessing a training image; 2), performing pre-training on a deep convolutional neural network by means of the preprocessed image, using a softmax loss as a loss function and introducing key point pooling technology; 3), inputting all training images into a pre-trained model, and calculating the initial kind center of each kind; 4), performing fine adjustment on the pre-trained model by means of the aggregate loss, aggregating the samples of the same kind to the kind center through iteratively updating a network parameter and the kind center, and simultaneously increasing distances between different kind centers, thereby learning robust discriminative face characteristic expression; and 5), in application, performing preprocessing on the input image, and respectively inputting the input image into the trained network model for extracting characteristic expression, and realizing face recognition through calculating similarity between different faces. The face recognition method can realize relatively high face recognition accuracy just through training small mass of data.