Forest Fire Smoke Identification Based On BPNN-DIOC Network

As the main body of renewable natural resources and terrestrial ecosystems on the earth,forests play an irreplaceable role in the history of human survival and development and become an indispensable material foundation for the development of the national economy.Forest fire is one of the most terrible disasters in forests.It not only burns forests and injures forest animals,but also reduces forest regeneration capacity,causes soil poverty,destroys forest conservation water sources,pollutes the air.Consequently,the ecological environment on the earth is over balance.In short,forest fires can bring many kinds of harms and are difficult to extinguish,so it is especially important to prevent them from happening.Smoke is a sign of forest fire and video surveillance has become an important means of monitoring forest fires.However,video surveillance systems are still dominated by manual monitoring and supplemented by semi-automatic recognition in recently application scenarios.The automatic monitoring technology is still not mature because of its low recognition rate and high false alarm rate.As one of the automatic monitoring technology,artificial neural network has been a hotspot in this field.However,because of the small amount of videoimage data,the selection of complex structures of neural networks is limited.Because of its simple structure,small training data set and fast convergence speed,BP neural network is suitable for automatic smoke identification and alarm.However,the BP neural network is a highly nonlinear network,so the influence of linear factors on target recognition is not considered in the structure,which leads to the low recognition rate of BP neural network in many applications.Therefore,the BPNN-DIOC neural network is proposed,which is used on the study on smoke identification in forest fire.The research results are as follows:1)Firstly,inherited the advantage of BP neural network,the direct mapping of the input layer to the output layer is added to BP neural network to form a neural network combining linear and nonlinear model,namely BPNN-DIOC.Then,aiming at solving the problem that neural network is easy to overfit and the generalization is not good,the classical LM algorithm is improved.To be specific,a regularization term is added after the loss function to improve the generalization of the network.Finally,based on statistical methods,the effectiveness of the network is evaluated by using different data sets in UCI database.It is found that the BPNN-DIOC network overall improves the classification performance and reduces the number of the hidden layer neurons in the network.2)The moving object extraction of video smoke image is carried out.The general moving objects are extracted by optical flow method,inter-framedifference method and background difference method.This paper compares the advantages and disadvantages of the commonly used method of moving target extraction,and improves the background difference method,namely the Gaussian mixture model method.According to the characteristics of smog,and finally adopts the Gaussian mixture model method to extract the moving target of smoke image.3)Analysis of forest fire smoke attributes and characteristics.Firstly,according to the characteristics of smoke,this paper analyzes the color of smoke based on the RGB color space,and the objects with strong interference are excluded.Then,the characteristics of smoke convexity,area growth of smoke movement,shape irregularity and background blurring characteristics are analyzed.The features of forest fire smoke images are extracted to prepare for the data input of BPNN-DIOC network.4)Experimental verification.The BPNN-DIOC forest fire smoke identification network is designed according to the smoke example,and recognition effect of BP neural network is compared with that of BPNN-DIOC network.The classification and identification of smoke on the BPNN-DIOC network can be used for early warning when the smoke appears in the video,and has strong robustness.The alarm accuracy can reach 94.80%,which has certain availability.