Non-sufficient training samples cannot comprehensively convey the possible changes such as illumination, expression and gesture, so it is hard to improve the accuracy of face recognition. To overcome the problem, Yong proposed a method that exploits the symmetry of the face to generate new samples and perform face recognition. The new training samples really reflect some possible appearance of the face. However, it usually gets bad symmetrical face samples based mirror image with the changes of facial poses, which may affect the accuracy of recognition. The SVD has advantages of stability and shift in-variance, which can ensure the rate of recognition in the case of small changes of face images. To ease the shortage of the above method, the authors improved it by generating‘symmetrical face’training samples based SVD and mirror image, respectively. The experimental results in ORL, PEFET and UMIST databases show that the improved method outperforms the effect of Yong's method.