Abstract: Under undetermined and sophisticated environment,it is an unresolved puzzle for early flame detection based on image.Therefore,for the improvement in detection efficiency,this paper not only introduced the theory of histogram equalization to this field,but also brought in the SIFT algorithm to image process of determining multi-scale features of the flame pictures,identifying flame extreme point as well as feature matching.We regarded fractal dimension as one of flame features.Since flame is the anomalous value mostly and there are some advantages in one class classifier,such as low cost,easy to obtain features and high precision,one class classifier is used to identify the flame.The experiment proves in this research that there is excellent rate of true positive and false positive under the circumstances of close quarters and bright light,in addition,there is high detection rate of flame and low false alarm rate under the dim light condition.

Key words: SIFT algorithm,One class classifier,Fractal,Histogram equalization,Flame detection,Flame sharp angle