Early Fire Automatic Detection And Recognition In Large Space Based On Digital Image Processing Technology

With the fast development of economy and society in our country, fire crashes in variouslarge space buildings are increasing. Because of the interferences of various environmentalfactors, the traditional fire detection and recognition technologies are difficult to detect andrecognize the early fire precisely when applied to large space occasion. The automatic earlyfire detection and recognition technology in large space based on digital image processinguses the surveillance video images as medium, and makes use of certain digital imageprocessing algorithm to detect and segment suspicious region, extracts and analyzes thefeatures of suspicious region, and accomplishes intelligent recognition by combining objectrecognition technology. Compared with traditional fire detection technology, this kind oftechnology is able to obtain more rich and intuitive fire information, and through itsnon-contact detection style, the technology shows unparalleled superiority in the sides ofeffectiveness, robustness and sensitivity, is very propitious to early fire detection andrecognition in large space.This paper takes early fire flame in large space as the main detecting object, fully makesuse of the visual features of early fire flame images to carry out research of the automaticearly fire detection and recognition technology in large space based on digital imageprocessing, and presents a set of automatically detecting and recognizing algorithm of flamesurveillance images.Firstly, in order to improve the quality of original surveillance video images, thusguarantee the follow-up detection and recognition of flame region to be accomplishedsmoothly, we carry out the research of surveillance video image preprocessing technology.The emphasis is to achieve contrast enhancement and smoothing filtering to video imagesequence, thus remove the noises and distortions that mixed into the images, prominent theuseful information of images, thus improve quality of images.Secondly, based on the surveillance video image sequence of fire flame, and berepresented by thresholding, we firstly analyze merits and drawbacks of the imagesegmentation technology applied to video flame detection which is based on single spatialdomain iamge processing. In allusion to the limitation of traditional image segmentationtechnology, this paper makes use of the feature that the intensity, color and region shape ofearly fire flame are in constant change, take the flame as a kind of special motion object,propose that put the motion object detection technology into the application of flame imagedetection, and build the self-adaptive gaussian mixed background model of surveillance video image pixels. This kind of model is able to update its parameters according to the change ofsurveillance enviroment, thus to guarantee effectiveness and robustness of motion objectdetection. In order to remove the interference of other kinds of motion object, thus make thedetection result more precise, we implement flame pixel color distinguishing andtime-frequency flickering analysis the detected motion object region, thus improve accuracyof flame region detecting algorithm.Thirdly, based on the detection result of flame video image, implement feature analysisand rescription to the segmented flame image, extract a variety of quantitative featuredescriptors of alternative flame region. This paper starts from the three aspects of colorfeature, texture feature and shape feature to describe the stationary visual feature; at the sametime introduces two kinds of descriptors of area changing rate and shape similarity to describethe dynamic visual feature, as the sullpement of the stationary visual feature of flame, thusimprove the effectiveness and robustness of flame recognition system.Fourthly, combining with object recognition technology, select the eight kinds of featuredescriptors of first order color moment, contrast, correlation, energy, circularity, area changingrate and shape similarity of flame image region, to form the feature vector fire pattern; takethe BP neural network as carrier, construct fire object recognition system, and select a certainamount of fire flame feature samples to train the BP neural network. The simulationexperiment results indicate that the trained BP neural network has nice recognition effect tothe visual characteristic signal of fire flame image, and can be used into the development andapplication of intelligent fire video surveillance system.