The conventional Maximum Margin Criterion (MMC) is sensitive to noise. To address this problem, a Robust Maximum Margin Criterion (RMMC) based on complex kernel is proposed. The samples are first mapped to the Complex Reproducing Kernel Hibert Spaces (CRKHS), and then the MMC is conducted in CRKHS. Besides, an efficient algorithm for implementing MMC is also proposed in this paper. The experimental results shows that RMMC is robust to the noise images and its recognition rates are higher.