Abstract: To accurately locate the fire source and achieve early warning of fire, a real-time fire warning monitoring method based on human visual attention mechanism was developed. Firstly, the brightness and color features of each frame of the video sequence were extracted according to image opponent theory. Secondly, pixel-level saliency detection algorithm was applied to construct a multi-scale spatial Gaussian pyramid which describes feature information. Then, the static saliency map was generated by merging center-neighbor contrast pyramid through the cross-scale feature addition method. Finally, the saliency map obtained from FMF algorithm was used as the dynamic frame difference based on the dynamic frame difference method to find the region of flame on video frames, and the proposed approach was compared with 6 representative algorithms in terms of 4 performance criteria on public datasets. The results show that FMF algorithm demonstrates stronger robustness in describing multi-scale spatial feature information through the saliency analysis method, and with obvious advantages in accuracy and missed rate compared with other algorithms, it can accurately identify and locate the flame so as to prevent the occurrence of fire accidents.

Key words: flame detection, opponent theory, saliency detection, fusion of multi-feature (FMF), dynamic frame difference method