Research On Fire Detection And Recognition In Coal Mine Based On Image Features

Fire is a kind of tremendous disaster in coal mine. It does great harm to people’s health, natural sources and influences safe production of coal mine. With the development of science and technology, automatic fire detection has become a significant resort for fire monitoring and prevention. Video-Image fire detection performs well in detection range, response time and cost, at the same time it is not easily subject to surroundings. Combined with computer technology, it can provide more intuitive and abundant fire information. Therefore, the research has great guide meaning to safe production of coal mine. Coal fire detection is composed of three parts: image preprocessing, feature extraction and image recognition. According to the three parts, main research work completed is as follows:In view of low illumination and uneven lighting in coal mine, a fuzzy enhancement algorithm is proposed and compared with traditional methods. A new fuzzy membership function is constructed. Threshold value is searched by a modified and rapid OSTU algorithm, and it can be obtained by self adaptability selection. Thus the pixel information which tends to be lost in low gray region is reduced and the image quality is improved. This algorithm improves the detailed information of the image, and increases the computation speed.In light of the large noise and low contrast in image, a new image segmentation algorithm based on two dimensional maximum fuzzy partition entropy is proposed. Fuzzy membership function is employed for subtle partition between object and background. Not only the present gray and spatial neighboring information is applied, but also the fuzziness of image is taken into consideration. It overcomes the disadvantage that the threshold is determined by manual work. Particle Swarm Optimization is applied to optimize fuzzy entropy function and increases the computation speed. The approach can meet real-time requirement.According to the flame image features in early stage, feature extraction of flame is investigated in detail. Combined with the static features, dynamic features and texture features, the candidate region and pixel are extracted and computed. In view of feature selecting for fire images, a multi-feature fusion technology is presented here. Redundant information of the original feature sets is removed; Finally, the parameters for classification on fire image are determined and are taken as recognition criteria between fire and inferences.After image features are extracted, fire detection model based on multi-feature fusion is constructed. BP neural network is suitable to deal with imperfect and fuzzy information, and Support Vector Machine(SVM) has advantage on small samples, non-linear and high dimension area, so BP and SVM are applied to fire image recognition, respectively. Objectives that each criterion is made best of can be obtained, and the limitation that single feature is acted as criterion is avoided. Based on fast leave one out method, the Fletcher-Reeves(FR) conjugate gradient method is employed for optimizing Least Square Support Vector Machine(LS-SVM) hyper parameters, and the FR-LSSVM model is established. Finally, experiments are made by BP neural network, FR-LSSVM, LS-SVM, and SVM, respectively. The results demonstrate that FR-LSSVM has faster computation speed and higher detection rate than the three other methods.According to the smoke image features, a real-time smoke detection algorithm is proposed. Firstly, the moving region is detected based on mixture Gaussian model from image sequence, and the moving pixel is extracted. According to the diffusing behavior and area increasing feature of smoke, the non-smoke object is removed. The smoke flicker frequency is investigated. After dynamic behavior and texture features are extracted and fused, SVM classifier is established for smoke detection.