Research Of Dynamic Texture Classification Based On Deep Learning And The Application In Forest Fire Prevention

The occurrence of forest fires brings huge economic losses to humans.Therefore,detecting and fighting fires timely are very necessary.The smoke is a form of dynamic texture,the dynamic texture feature extraction has a very important effect on the smoke recognition.Therefore,the dynamic texture classification method has great practical significance.In this paper,deep learning is used for dynamic texture feature extraction,and successfully applied in the classification of dynamic scenes.This thesis introduces the existing classification methods of dynamic texture and three kinds of typical deep learning network structure,namely Deep Belief Network model?Stacked Denoising Auto-encoder Network model and Convolutional Neural Network model.And the thesis proposes the main work in the following areas:Firstly,a slow feature analysis is proposed based on Stacked Denoising Auto-encoder Network model for dynamic texture feature extraction.The essence of slow feature analysis is to extract signal features from changing slowly input signals,however,nonlinear expansion's Limitations of input signal makes the output dimension larger,Stacked Denoising Auto-encoder Network model has a good effect on dimensionality reduction.Experimental results show that this method compared with existing methods,greatly improves the accuracy of dynamic texture classification.Secondly,thesis proposes the 3D convolution neural network on the dynamic texture classification based on the two-dimensional convolution neural network.The original 2D Convolutional Neural Network have a good effect on two-dimensional image feature extraction,but it does not use in video sequences.Convolution and down-sampling layers are alternately stacked to form a network structure.Experiments show that the network model can extract more advanced video texture features,and have a good effect on the dynamic scene classification.Based on the above characteristics of dynamic texture classification methods,thesis build and design a forest fire surveillance system in the Visual Studio 2010 environment,which realizes the function of real-time video browsing,fireworks identification,fire alarming,pan-and-tilt camera control and others.The experimental results show that the system can complete the intended target recognition stably,meet the basic requirements of fires recognition.