Component Separation Algorithm For Image Smoke Recognition

Compared with the smoke detection technology based on sensor principles,smoke detection based on the video technology has been insensitive to environmental factors,high responsed speed and visual detection results.It is an important means to realize a early warning of fire.At the beginning of the fire accident,there has been usually developed a great deal of smoke.The main research work of this thesis is to determine whether there is smoke generated by monitoring the scene.Existing video-based smoke detection methods often rely on the visual features extracted directly from the original frames,ignoring a semi-transparent feature of the smoke,the features contains the background information and smoke information in two parts,which can not effectively describe the feature of the smoke and can not guarantee the accuracy of the algorithm.Faced with this problem,this thesis assumes that an image is composed of a linear blending of a smoke component and a background image from the point of view of imaging principle.This thesis presents a smoke linear expression model and an optimization problem.In order to solve the problem,the component separation algorithm is proposed,and the component separation algorithm to separating the smoke component from a current image and extracting from the smoke component for smoke recognition.According to the correlation between adjacent pixels and pixels,the algorithm has constructed Local Smoothness model using the patch as the calculating unit.Meanwhile,from the perspective of overall textural configuration,pure smoke image are likely to lie in a low-dimensional subspace,and principal component analysis(PCA)can be used to determine pure smoke image subspace with describing the pure smoke image.Therefore,the principal component model is constructed in this thesis.There are two models to separating the smoke component from an image for pure smoke components.Then extracting the texture feature of smoke components by the LBP operator.Finally,it uses support vector machine(SVM)classifier to determine whether there is smoke or not,so as to realize smoke recognition.Experimental results on synthesized images and real video data have shown that the proposed approach can effectively separate the smoke component and compared the experimental results with Toreyin and Tian in the detection accuracy aspect.Experimental results also show that this algorithm can effectively identify the smoke in indoor and outdoor,complex background,and positive detection rate is above 93%,the false detection rate and missed detection rate are below 4%.This thesis is applicable for the detection of different types of smoke,such as full-covered heavy smoke images,full-covered light smoke images,partially covered more than 50%of the smoke images,partially coverage is less than 50%of the smoke images.The average accuracy of this algorithm is higher than smoke recognition algorithms of Toreyin and Tian.In this thesis,the average detection accuracy is 91.3%,the false detection rate is about 5.4%,and the missed detection rate is about 3.3%...