Reaearch On Extraction Of Fusion Features And Classification Algorithms Of Flame And Smog

In the course of fire prevention, fire automatic detection and recognition is very important. The sensing device based traditional fire detection technology has many deficiencies. Not only vulnerable to interference from environment, but also lack of recording of situation when the fire broke out which brought difficulties to subsequent investigation. In this case, the fire detection based on video image come into the sight of domestic and foreign scholars. Video image recognition based fire detection can not only increase the speed, but for a wide range of fire monitoring also has a great advantage, you can also record the situation when the fire broke out. This paper probes into the preprocessing of flame and smog image,the feature seletion and classification algorithms of flame and smog regions.In the feature selection part, this paper introduces the covariance characteristics based on flame and smog area. It presents a method of producing optimal classification feature of the sequence based on degree-of-contribution for optimal classification to these candidate color and moving subsets by relief analyzing and PCA transformation. It also validates contribution ratio of discrimination of all selected colors and moving features by covariance matrix. It probes characteristic description by HOF and HOG as flame and smog’s spatial-temporal features. It proposes an analysis way of histograms of oriented optical flow in different channels. It probes how to describe construct fire region’ s HOFHOG visual dictionary by k-means.In the classification algorithms part, there are 2 classification algorithms, support vectors machine and random forest. A new algorithm that combined SVM with KNN is presented and it comes into being a classifier. It has probed parameter selection and performance analysis for random decision forest classifier training based on feature subsets via relief feature selection.It detects flame and smog region simultaneously and submit a more logical alarm judgment by decision tree forest voting according to detected region’s spatial-temporal distribution and relations.At last, this paper integrates the image processing methods introduced before and the results of the experiments, summarize that the flame and smog detection system based on random decision forest classifier has stable higher accuracy. A detection softwart, which suits for complex environment, of flame and smog is developed based on therories of features extraction and classification algorithms mentioned before.