Research On Fire Image Detection Technology Based On Flame Features

Nowadays, as one of the severe disasters that occur frequently, fire leads great threats to the safety of human life and property. So how to recognize the fire accurately and efficiently has become an urgent issue that needs to be addressed. Traditional technology for fire detection is vulnerable to environmental impacts and has a high rate of false alarms. By contrast, fire detection technology based on video images, which identifies the fire by extracting the features of flame or smoke in images, is excellent in response time,recognition rate and environment adaptability. Because of these advantages, video-based fire detection is increasingly becoming a research hotspot in recent years.In this thesis, fire detection technology based on flame features is expansively analyzed,and the main research content and innovations can be summarized as follows:Firstly, the current status of research and development of fire image detection technology along with its principle and characteristic are introduced in detail based on a great deal of domestic and foreign literature.Secondly, on the basis of foreign fire data set, a segmentation method that can identify and enlarge suspected flame regions using statistical model based on RGB color space is proposed in the thesis, and the recognition accuracy of this approach is compared with that of other color models.Thirdly, several dynamic and hierarchical features of flame images are extracted and analyzed emphatically, which includes the change rate of flame areas, the flame similarity,the flame flicker frequency, the statistics of gray difference, and the coefficient of gray difference variation, and so on. In addition, by comparing the differences of these relevant parameters between fire and non-fire videos, the above features are shown to be effective and reasonable for the recognition of flames.Finally, the five selected characteristic parameters are processed and fused by a BP(Back Propagation) neural network for fire detection. Testing results show that the flame recognition algorithm presented in this thesis is robust and efficient, and it can accurately recognize the fire flames in different scenarios, and eliminate the highlighting interferences and flames that are under control or in safe conditions.