Flame Detection In Videos Based On The Analysis Of Visual Saliency

In current society, fire is one of the most frequent disasters. It not only threatens people’slife and property seriously, but also influences the economy and the development of thesociety. So it is very significant for fire prevention. One important way to fire prevention isfire detection real-time. But the traditional fire detectors using light, heat or smoke have manyshortcomings, such as small detection range, slow response speed, high false detection rateand missing detection rate, etc. With the rapid development of image processing and patternrecognition based on computer technology, video fire detection based on computer visiontechnology shows significant advantages. Video fire detection has become a research hotspot.In this paper, the visual characteristics of fire were elaborated in details, and theadvantages and disadvantages of existing video fire detection technology were introduced andanalyzed. Then video fire detection based on visual saliency was studied deeply.Because the selective attention mechanism of human visual system can not only locatingsalient objects in an image quickly, but also doesn’t need to be trained when people look atimages. So the saliency method was introduced to the field of computer vision, and thetop-down mechanism of visual attention was integrated to implement the measurement of thesaliency of flame’s key characteristics. Finally the video fire detection model of space-timevisual selective attention mechanism was constructed.Brightness, color, texture and flicker are all of salient features of flame. Firstly, thesaliency of flame’s brightness and color were defined. Because the popular flame colormodels in RGB space cannot extract bright yellow parts in the center of flame areas, aneffective color formula which can extract the bright yellow parts completely was proposedand it filled the holes of the flame areas to some extent. Finally, the small holes formed bybright white parts of the center of flame areas were filled by morphological operation.To avoid the texture feature vectores of the train samples or test samples entered intotime-consuming Support Vector Machine (SVM) to be trained and classified directly, themethod of Principal Component Analysis (PCA) was used to reduce the dimensions of LocalBinary Pattern (LBP) texture feature vectors of flame areas in the train samples. A standardLBP feature vector was got. Then the saliency of flame’s texture was obtained after thedistances between LBP feature vectors of the suspected flame areas and the standard featurevector were calculated.Because false detections by using the ordinary methods of flame moving targetsextraction and time-consuming for establishing a complex model about flicker for suspectedpixels, the novel saliency of the motion of flame was defined by two ways, before which themethod of accumulative difference was adopted, and flame flicker frequency and the framerate of videos were considered. The first way to do this was to use the brightness change ofsuspected flame pixels between video image sequences. The other way was to compute thenumber of high frequency energy zero-crossing of suspected flame pixels between videoimage sequences on which the wavelet transform was performed before.Two methods were proposed to get final flame saliency map. The first was to use the linear weighted method to fuse the flame features of brightness, color, texture and motionaccording to brightness change.The other one was to use quaternion discrete cosine transformto integrate the flame features of another two color, texture and motion which uses the numberof zero-crossing. The former is simple and effective, while the latter improves the saliency offlame targets furtherly.In this paper, all the13fire videos and2non-fire videos of Bilkent university flamevideo library were selected for experiment in addtion to4non-fire videos obtained from theInternet. The experimental results showed that the proposed model has the characteristics ofhigh accuracy and strong robustness compared to other popular models, because of theaccurate and stable static and dynamic characteristics in this paper.