Fire prevention has gradually dominated by video surveillance. There are many various algorithms of video flame, but most of them require lots of video flame samples to do the training for the final classification. The results often fail to achieve a high detection rate if flame samples are not enough. So a new method that uses the super pixel segmentation and the flash frequency for recognition is proposed in this paper. In the Lab color space, the cencroids of each region are represented as a super pixel after some approximate homogeneous regions are segmented in the flame picture based on super pixel segmentation. The candidate regions are extracted based on the RGB and Lab color features according to some rules. At last whether the candidate region is a flame is determined by the flash frequency characteristics. Experimental result presents good performance with high detection rate in the case of small samples.