Forest Fire Recognition Based On Convolutional Neural Network

Forest resources are a precious asset to mankind,maintaining the balance of the ecosystem,and in the event of a forest fire,the impact on forest resources is enormous.The hazard.Therefore,it is important to prevent forest fires from occurring.Traditional forest fire identification algorithms are sensor-based,which are greatly influenced by the environment and are not applicable to open areas.Most of the image-based forest fire identification algorithms rely on manually extracted features,and the selection of features relies on the experience of researchers.Deep learning takes a large number of samples as input and automatically exploits the features of multi-layered neural nodes,which in turn leads to better generalization capabilities of the model.Therefore,a convolutional neural network based forest fire identification algorithm is proposed in this paper which is as follows.1.Summarize the existing forest fire recognition methods at home and abroad,mainly including sensor-based forest fire recognition algorithms,traditional manual feature-based image processing forest fire recognition algorithms,and deep neural network-based image processing forest fire recognition algorithms.And analyze the advantages and disadvantages of the above methods..2.To address the lack of forest fire image samples and the poor generalizability of traditional image-based forest fire recognition methods,this paper a method for forest fire identification based on convolutional feature fusion in a complex context is proposed.The method consists of two parts:augmentation of forest fire samples based on the CycleGAN technique in the field of image style migration,and augmentation of forest fire samples based on convolutional feature fusion.Feature Fusion for Forest Fire Identification.In the first part,the lack of forest fire image samples is addressed by applying CycleGAN based on the image style migration domain.Techniques to generate forest fire images in complex backgrounds and expand forest fire samples.In the second part,the forest fire images are first pre-processed and then feature-learning using deep neural networks will be used to extend deep it is fused with the shallow convolutional features for final forest fire identification.The experimental results show that the method can achieve a better recognition effect.3.In view of the fuzzy shape of forest smoke,the obscure color relative to the background,and the diverse appearance characteristics,an Improved Multi-Scale Convolutional Neural Network(IMSCNN),which was first designed as a convolutional neural network a kind of basic block,basic block consists of several parallel convolutional layers with the same number of convolutional kernels of different sizes,each convolutional layer is followed by a batch normalization to normalize the output of the convolutional layer.Second,the output of the normalized convolutional layer is summed and the summed result is activated.Finally,a shortcut connection is used to fuse the feature maps of basic blocks of different sizes,and to obtain missing details from the feature maps of larger basic blocks.Information to improve the characterization of features.In order to compress the size of the recognition model and improve the recognition speed of the algorithm,the large convolutional kernels in the basic block are decomposed into multiple small convolutional kernels,and the asymmetric decomposition technique was further employed.It is experimentally demonstrated that the method achieves better smoke recognition compared to existing methods.