The Fusion Of Image Separation And Characteristic Analysis Of Smoke Detection Algorithm

Fire is a kind of natural disasters with high sudden and strong destructive power.Using fire carelessly will also give rise to a serious threat to people's life and property in production and living.Smoke appears before the flame in initial stage of fire.Smoke detection can win over more precious time for fire remedy.Therefore,smoke detection has become an urgent issue in nowadays society.At present,the video smoke detection becomes a new research direction because of its high efficiency,the advantages of non-contact and embeddablity.In order to improve the efficiency and reliability of video smoke detection system,the paper concentrated on three stage of video smoke detection technology and put forward the fusion of image separation and the characteristics analysis of smoke detection algorithm.The three stage are respectively the target foreground detection stage,smoke feature extraction stage and smoke recognition stage.(1)The foreground object detection:As for the foreground object detection,the Smoke detection segmentation algorithm with the fusion of ICA and GBVS was proposed,it improved the large sample size and long sample time span of traditional gaussian mixture model.Firstly,the algorithm got preliminary smoke foreground by using ICA preliminary separation model in smoke separation stage.Secondly,it extractedthe multi-channel and multi-scale characteristics of image through GBVS to get saliency smoke foreground area.Finally,it identified the smoke by histogram matching using the characteristic of color and texture.The experimental results show that the algorithm effectively reduces the scope of the smoke foreground area by combining ICA and GBVS in the smoke foreground separation stage.The smoke foreground area extracted by this way is not only small but also concentrated.The ROC curve shows that the algorithm performs excellent as a whole.(2)smoke feature extraction:As for smoke feature extraction,smoke detection algorithm based on multi-dimensional dynamic texture analysis was proposed.Traditional linear dynamic system(LDS)of video smoke detection only used brightness values as image information,it ignored the other information such as RGB values of video and its calculated amount was large due to intensive sampling.The algorithm first got smoke candidate area by preprocessing of smoke color filtrate and background subtract tion,it avoided the dense sampling and reduced the amount of calculation.Secondly,it added RGB and HOG features to the multi-dimensional image block,so it increased the dimension of image block.Finally it analyzed the dynamic texture characteristics of smoke video based on the decomposition of high order of multi-dimensional image data.Due to the sliding time window,it can determine the location of the images where smoke occur and the specific time when smoke occur.The experiment used detect rateas the evaluation index and it used the method of multiple comparison,the results show that the algorithm improves the reliability of the dynamic texture feature analysis,additionally its amount of calculation is small.(3)smoke recognition:As for smoke recognition,smoke detection algorithm based on fuzzy clustering combining generalized entropy with neural network was proposed.The algorithm firstly got smoke foreground by using the Smoke segmentation algorithm with the fusion of ICA and GBVS,and then extracted multi-dimensional dynamic texture feature based on higher order linear dynamic systems(h-LDS).Finally,it used a fuzzy clustering algorithm with generalized entropy based on neural network to train and classification.The experimental results show that smoke detection algorithm based on fuzzy clustering combining generalized entropy with neural network has higher accuracy and its training error is small.