Flame detection algorithm is one of the key technologies for power safety early warning. In order to improve the accuracy of flame detection, this paper proposes an algorithm based on word bag model. Firstly, the shape feature of flame is extracted according to the growth of flame area between consecutive frames. Then, the flicker frequency of flame is analyzed by the number of brightness jump of flame pixels in a period of time. At the same time, a new word bag model is proposed to describe the flame motion characteristics. Finally, the validity of these features as the criteria of flame recognition is demonstrated by experiments. The simulation results show that the flame recognition algorithm proposed in this paper can recognize the flame effectively.