Research On Pixel Annotation And Real-time Detection For Forest Fire Images

It is destructive for forest fire,which not only causes huge economic losses,affects the ecological environment,but also seriously threatens the safety of human life.In recent years,the frequent change of global climate has aggravated the frequency and harm of forest fire disasters.Therefore,the countries all over the world have devoted heavy investments,e.g.,human recourse and labour,to develop forest fire monitoring systems.Among them monitoring systems,the visible light-based monitoring system wins popular beacause of the characterists of lower-cost,real-time,and larger view area.According to the data of the Ministry of Emergency Management of the P.R.China,more than 97% of the identified forest fires were caused by human factors from 2010 to2019.As far as the current monitoring system is concerned,however,the main difficulties are still focused on the recognition of "fire",which is caused by the following reasons: 1)the uncertainty of research objectives,which is manifested in the uncertainty of the color and shape of the flame,the complex and changeable forest scene,the serious mismatch between the flame target and the non-flame target,etc.;2)the research methods,most of them still equate the problem of forest fire recognition with the problem of binary classification,and ignore the preconditions of learning methods,such as independently identically distribution;3)there is no fair standard of forest fire dataset,especially the data with accurate class labels,and the annotation itself is not only timeconsuming and labor-intensive,but also a huge workload.In order to solve or alleviate the above difficulties,this paper completes the following work from the perspective of machine learning.1.The existing work mainly focuses on the selection of color space.This paper discusses the separability of forest fire pixels in RGB,YCb Cr,HSV and HSI color space.That is to say,three classical classifiers(linear or non-linear)are used to compare the classification accuracy of fire pixels and non fire pixels on several marked forest fire images.In order to avoid the under learning or over learning problem of the classifiers,the experimental results are as follows The experimental results are as follows: 1)the separability in RGB and YCb Cr spaces is equivalent;2)although the transformation between RGB and HSV or HSI spaces is nonlinear,the change of separability can not be known in theory,but the experimental results show that the classification accuracy in RGB space is higher than that in HSV and HSI spaces,so it is suggested to carry out the experiment directly in RGB space.2.In the research process of question 1,it is very difficult to obtain the ground truth.In order to overcome this problem,this paper proposes a pixel level automatic image annotation algorithm directly in RGB space,namely KNN annotation method based on KD tree,in order to realize automatic annotation of forest fire image or video.The details are as follows: 1)breaking through the conventional idea of regarding forest fire problem as two classification,we plan to use multi classification method for classifier design;2)easing the restriction similar to independent and identically distributed condition,using distribution free KNN to design classifier;3)considering the subjectivity of category number selection,the algorithm adopts interactive design,and the number of categories is selected by the user In order to avoid the repeated training problem caused by the change of training set,KD tree is used to store the training samples;4)in order to select the pixels in the region of interest conveniently,the empirical bounding box method(mostly rectangular box)is extended,and a fast selection algorithm based on convex hull is proposed.Compared with the existing rule method or binary classification method,analysis and experiments verify that the method has high fire detection rate and low false warning rate,which can be used as ground truth of forest fire image.3.In order to solve the problem of slow recognition speed caused by pixel by pixel scanning,this paper proposes a fast block by block detection method,which uses super pixel segmentation method,such as SLIC(simple linear iterative)In this way,the sample size of block pattern can be greatly reduced compared with that of pixel vector pattern.In this paper,the mean value,median value and their combination value are selected as the representative points of the block,and the experiment shows that the combination value performs best.On the premise of achieving ideal flame detection rate,the algorithm performs well in real-time,and meets the real-time requirements of forest fire monitoring.