ing at the problems that the traditional normalized exponential loss (Softmax) function lacks the ability to distinguish features and facial features cannot be learned discriminatively, an aggregation and discriminant multitask learning algorithm is proposed. First, the multitask cascaded convolutional neural network is used to detect and align the face of the target images, eliminating the images that are not related with the face recognition region. Then, the deep convolutional neural network is used to extract the aligned facial features. The aggregation and discriminant multitask learning algorithm is used to decompose the vectors of extracted facial feature into learning intra-class features and discriminating inter-class identities, which strengthens the constraint of intra-class features and improves the separability of inter-class features. Finally, the nearest neighbor classifier and ten-fold cross-validation method are used for face recognition and verification respectively. Experimental results show that the verification accuracy of the proposed algorithm in LFW face database can reach 99.68%. The proposed algorithm improves the performance of face recognition, and has good robustness in illumination, pose, expression, and age change test. It can be effectively used in face recognition engineering practice.