Study of Early Fire Detection Technology Based on Image Features
|School||Qingdao Technological University|
|Course||Heating,Gas Supply, Ventilation and Air Conditioning Engineering|
|Keywords||fire image detection image segmentation feature extraction neural network identification models|
With the accelerating development of urbanization and the rapid increase in urban population, there are more and more large-space and high-rise constructions. The buildings have the characteristics of complex internal structure and high population density. Once the fire broke out, it spread quickly and hard to put out, in this case, the evacuation is also difficult. So it is crucial to take effective precaution measures and to detect and control the fire as soon as possible. The fire image detection technology, combined with the advantages of visual images and the high-speed processing of computer, is a novel, effective early fire detection technology. As the technology can detect large space fire promptly and accurately, much attention has been focused on it by scientific researchers and engineering technicians.The key technologies of fire image detection, such as the segmentation, feature extraction and identification of fire images have been studied. According to the respective characteristics of fire smoke and flame, and the analysis of their determinacy and randomicity, we propose to obtain the smoke segmentation threshold by optimum threshold value and to get the flame segmentation threshold by experiment, and then segment fire smoke and flame separately by background subtraction method. The experimental result indicated that this method has a good segmentation accuracy and speed. We use Multiresolution Analysis (MRA) and the secondary statistic of Gray Level Co-occurrence Matrix (GLCM) to innovatively analyze and extract the fire smoke’s textural features, and have designed four kinds of feature vector selection modes. The result indicated that the feature vector matrix constituted of secondary statistic of the original image and the energy-dominant wavelet packet nodes have the highest fire smoke identification accuracy. We analyze and quantify the rules of early flame’s area growth, edge vibration and centroid invariability, and have compiled the corresponding feature extraction program. In view of the features of fire smoke and flame images, we have separately designed the fire smoke and the flame neural network identification models. The experimental result shows that the neural network models can effectively recognize the fire smoke and the flame. In the end of this dissertation, we analyze the outputs of the fire smoke and the flame neural network identification models, and combined the two models so as to effectively improve the accuracy and the stability of fire recognition.