The small sample problem has been a major challenge for face recognition applications for a long time. Concerning the insufficient sample problem in the face recognition process, we propose a face recognition algorithm based on QR decomposition and reconstruction of virtual training samples. We use partial Q and R information to construct virtual samples which have certain difference from the original face image, thus more effective characteristics of the face image that likely change with emotions and illumination are increased, and the training sample sets are expanded. Then we perform weighting fusion on the collaborative representations of the original samples and virtual samples, select the optimal weight combination, adjust the proportions of the original sample and virtual sample which can impact on the results, and obtain the correct recognition rate. Experiments on the ORL, FERET and AR face data sets show that the proposed algorithm can achieve a higher recognition accuracy.