Abstract: Recently, the sparse representation based classification has been proved to be superior to conventional Face Recognition (FR) methods. Though the sparse representation based classification can achieve striking recognition performance, the classification accuracy still has room to grow. To improve the accuracy of FR, a kernel sparse representation method based on virtual samples (KSRVS) is devised here. It first generates virtual samples and then uses a kernel-induced distance to perform FR. Because more training samples can provide more useful information in representing the testing sample and the kernel-induced distance can select the training samples that are truly “near” to the testing sample, the final error ratios obtained using the KSRVS are lower than various sparse representation methods such as the sparse representation method based on virtual samples, kernel-based sparse representation method, two-step test sample sparse representation and the feature space-based representation method, collaborative representation based classification with regularized least square.