The Research On Forest Fire Flame Recognition Based On The Image Of The Video

Forest fire is an important natural worldwidedisasters, and more and moregovernments pay attention to them. Because forest fires have characteristics such asspreading fast, puting out the fire and rescuing be more difficult. So the problembecomes more and more urgent that how to prevent them effectivly as well as earlydetection of forest fires. In recent years, with the rapid development of visualizationdevices and digital image processing techniques, the researchers propose a newflame detection method based on the video image. The new flame detection methodbased on the video image integrates of computer vision, digital image processing,pattern recognition and signal processing and other aspects of technology.It is notlimitated by the environment and distance and can automatically identificate thevideo flames.The most important is it can play a role in the early warning of fire.This paper studies on forest fires flame video recognition method based onvideo images respectively from the detection of the flame zone, feature extractionand flame recognition:First, we can use histogram equalization and filtering on the video image andextract the suspected flame pixels with movement characteristics and colorcharacteristics of the flame image; Then we can use K-means clustering algorithmto extract the flame region;Last we use morphological methods to processe the flameregion and draw contour lines.Second we extract the static feature vector flame feature vectors correspondingto these regions and also calculate the texture information that includes energy,entropy, moment of inertia and related features including a rectangular shape,elongation, roundness, invariants moment1. Then according to the spread of flameand flash of dynamic characteristics,we can extract the five dynamic featuresincluding frame rate of change of the adjacent area, the rate of change of the heightof adjacent frames, sharp features, frequency characteristics, the movement of thecenter of mass.We can choose features include a rectangular shape, elongation,roundness, invariant moment1and the frame rate of change of the adjacent area, therate of change of the height of adjacent frame to generate the feature vectors, theexperimental results demonstrate the effectiveness of feature selection.Finally we identiy these feature vectors respectively through LVQ neuralnetwork and BP neural networks and support vector machines.Experiments cancompare the recognition performance and training speed in LVQ and BP neural network.For support vector machines, we want to study the recognition rateunder different kernel functions.Ultimately, we will find support vector machine isrelatively high recognition rate points through a comparison among the effects of thethree methods.