Research On Forest Fire Recognition Method Based On Deep Transfer Learning

In recent years,forest fires occur frequently.Therefore,the research on forest fire prevention technology is very necessary.At present,the application of computer vision and artificial intelligence technology in the field of forest fire prevention is still in the initial stage.In this paper,LBP algorithm,deep learning method and migration learning method based on confrontation are improved and innovated.An improved method based on local binary pattern(LBP)is proposed.Specifically,LBP algorithm is used to extract the LBP features of the forest fire smoke sample image,and then the image features of the training samples are used to build the feature dictionary.The reconstruction error of the test sample in the feature dictionary is obtained and compared with the set threshold value.When the reconstruction error is less than or equal to the threshold value,it is identified as forest fire smoke image.In order to test the influence of the number of samples in the feature dictionary on the recognition ability,experiments are set up to detect the change of recognition rate when the number of samples is different.Finally,the control experiment was carried out and VGGNET and RESNET were set as the control group.The recognition rate of the method proposed in this chapter is 6.11 and 5.61 percentage points higher than that of the control group.It can be seen from the above experiments that the number of training samples has a certain impact on the recognition accuracy,but when the number of samples reaches a certain order of magnitude,the impact on the accuracy will gradually weaken.In addition,compared with the mainstream deep learning network,it has a significant improvement in detection accuracy and detection speed.A method of forest fire smoke recognition based on 3D residual dense network is proposed.This method is improved on the basis of the three-dimensional convolution neural network.Compared with the traditional three-dimensional neural network,each convolution layer of the model is connected in a dense way to improve the data transmission capacitybetween the modules.At the same time,the model parameters are improved.In order to test the performance of the model in the identification of forest fire smoke,C3 d and 3D RESNET were selected as the control group to carry out the control experiment,and the accuracy rate and the missing detection rate were used as the basis for discrimination.Compared with the other two methods,the accuracy is 1 and 2 percentage points higher.The experimental results show that the model has better performance in different forest environments,and the improved model has a greater improvement in the extraction of sample feature information.A transfer learning method based on domain cooperation and domain antagonism is proposed.In this method,each layer of deep network is regarded as a module,and a layer of discriminator is added to each module.And define a loss function on each module.The main idea of this model is that the discriminator can’t distinguish the source domain from the target domain through adversary learning in the high-level modules.In the low-level module,make the source domain and target domain as similar as possible,so that the target domain can learn the edge information of the source domain well.The model is trained in this way.At last,vgg-16,RESNET and RESNET transfer learning were used as the control group in data set mivia and cifar-10 respectively.Compared with other methods,the accuracy of mivia data set was improved by 3.38,2.91 and 2.26 percentage points respectively.The accuracy of cifar-10 is 4.63,4.3 and 1.64 percentage points higher than other methods.It is proved that this method can improve the detection accuracy of forest fire smoke image.The method of this paper detects the model on two data sets respectively.The research goal of this paper is to improve the detection speed and accuracy of forest fire smoke and avoid the over fitting problem caused by insufficient data.Three methods are proposed for different problems,and their effectiveness and feasibility are verified by experiments.