Objectives  In order to solve the problems of the low accuracy and the slow speed of the existing video image flame detection methods, we propose a real-time video flame detection algorithm based on improved Yolo-v3 to achieve real-time and efficient detection of flames in the video.

  Methods  Firstly, in the feature extraction stage, the multi-scale detection network is improved. We add a new-scale feature and then improved the networks ability to learn the shallow information of the images by further integrating multi-scale features. Using this method, the accurate identification of small flame is achieved. Secondly, in the target detection stage, we use the improved K-means clustering algorithm to optimize the multi-scale prior frames, and make them adapt to the changing posture and shape of the flame. Finally, after detecting video flames based on improved Yolo-v3, we use the unique flicker characteristics of the flame to check the video again, and eliminate the false detection frame in the detection result. And in this method the detection accuracy is further improved.

  Results  In order to prove the effectiveness of our method, the video flames are detected from both accuracy and speed, and the results are compared with the advanced flame detection methods in recent years. The results show that the average accuracy rate of our method can reach 98.5%, the false detection rate is as low as 2.3%, and the average detection rate is 52 frames/s, so our method has better performance in terms of accuracy and speed.

  Conclusions  The effectiveness of this method is proved through multiple sets of experiments. Comparing with the existing flame detection methods, our method can be more effectively applied to video flame detection.