Research On Flame Detection Method Based On Infrared And Visual Information

Fire,as one of the major disasters in daily life,is extremely destructive and seriously threatens people's lives and property.How to effectively detect the occurrence of fire and avoid the spread of fire has been the focus of domestic and foreign research.Traditional flame detection technology has the shortcomings of single detection method,long response time,low reliability,and inability to provide effective location information for subsequent fire emergency treatment.In view of the above shortcomings,this thesis studies a new type of flame detector based on infrared and visual information.The new type of detector effectively combines infrared and image detection technologies,while using binocular positioning technology to obtain the flame position which provides effective position information for further fire treatment.The main research of this thesis maintains the following three aspects:Firstly,for the current image flame detection algorithms,when flame foreground is extracted,there are still shortcomings like incomplete flame contours and poor anti-interference.This thesis proposes a new flame foreground extraction algorithm,which combines the RGB color space model with the HSI color.The space model is fused for coarse segmentation,and then fine segmented by the maximum between-class variance method(Otsu).Using the two-color space fusion algorithm can extract a more complete flame profile,making the extracted flame profile less affected by interference.After the foreground image is obtained,flame feature extraction is performed,flame color features are extracted in the YCbCr color space,and flame texture features are extracted using a gray level co-occurrence matrix for the final flame judgment.At the same time,an improved probabilistic neural network(PNN)is proposed,which improves the smoothing factor in traditional PNN from a single fixed value to multivariate parameters.The conditional expectation maximization(ECM)algorithm is used to optimize the parameters of the smoothing factor in PNN.The extracted features the input of the training test in the improved PNN.Simulation results show that the algorithm has good anti-interference ability and can improve the accuracy of flame recognition.Secondly,for the defects that traditional flame detectors cannot provide the specific position of the flame,a binocular imaging model is established,and the position of the flame is detected by using a binocular camera to provide reliable position information for further fire suppression operations.First,we standardize the binocular camera with a checkerboard picture to obtain the camera's internal and external parameters;then perform distortion correction on the binocular flame image according to the parameters to obtain a line-aligned image.In image matching,we use the SURF algorithm and the FLANN algorithm to combine matching feature points,add prior region constraints and left-right alignment constraints during matching,it effectively reduces the complexity of matching calculations and improves the accuracy of matching;after image matching,calculations are performed according to the triangle similarity principle and calibration parameters,then specific flame position information is available.The experimental results show that the matching algorithm with constraints effectively improves the efficiency of image matching and improves the accuracy of matching.Finally,the ideal flame coordinates are obtained,and good positioning results are obtained.Finally,in view of the false-positive and high-latency of the traditional three-band infrared flame detector,a three-band infrared flame detection platform based on LASSO regression was built,and a specific recognition algorithm applied to the three-band infrared flame detector was proposed.The design of circuits and software programs was completed.In a complex industrial environment,the data collected by flame detectors are easily disturbed by environmental factors,so the extracted features are highly complex.This thesis uses the good prediction ability,coefficient compression ability and feature selection ability of LASSO regression to effectively improve the accuracy and sensitivity of the detector for flame recognition.At the same time,LASSO regression also has the characteristics of high efficiency,high prediction accuracy and strong interpretability.Experiments show that LASSO regression has improved the accuracy and real-time performance of traditional infrared flame recognition algorithms.