| Fire recognition is the most important part of the Fire Monitoring System, which has been the hot spot and difficult point for a long time. On the research of fire monitoring, although many algorithms have been proposed, it is still a challenging problem. In the early studies on fire detection, due to the limitations of the sample selection, not pay much attation to the real-time performance of system, etc, so it is limited to the actual application. What’s more, the technology development of fire detection in China has a great gap compared with developed countries.Based on the present situation of the fire detection technology at home and abroad, this paper mainly studies the methods of fire recognition based on image/video. This article mainly committed to fire image processing, flame segmentation and fire recognition. In the pretreatment of fire research we study the denoise methods of fire image. In flame segmentation, one thousand pieces of test images was statisticed in this method, then a new fire flame segmentation method was proposed based on Red-Channel. This method achieved good result compared with the traditional segmentation method. In video processing, through the combination of frame difference algorithm, the Anti-interference capacity has been strengthened. In terms of fire recognition, this paper adopted three fire recognition methods which based on fire abstract feature, and then we analyzed those three methods on the rate of accuracy recognition, the rate of false positives, non-response rates and the real-time performance of system. These three methods are the method of PCA and neural network, method of RBM and the method of ELM. These three methods are not need to extract dynamic or static characteristics of the flame, can forecast the fire accurately. The improved ELM method for fire recognition accuracy up to 95% above, the test time of 1800 samples is less than one second,this can satisfy the requirement of system real-time performance very well. |
PDF Full Text Request