Research On Forest Fire Smoke Recognition Method Based On Deep Learning

In order to reduce the economic and property losses caused by frequent forest fire accidents,it is crucial to monitor forest fires timely and accurately and contain them in the nascent state.Due to the complex background of forest fire images,the changing shape of smoke,and many interfering factors such as water mist and clouds similar to smoke,it is difficult for traditional image-based forest fire smoke recognition methods to extract smoke features accurately,and the models have low accuracy and poor generalization ability.The current deep learning-based smoke recognition method achieves smoke recognition by automatically learning image smoke information,but this method also has problems such as too few data samples,low recognition accuracy,and slow response speed.To address the above problems,an efficient monitoring and early warning method of forest fire smoke based on YOLOv7 is proposed.The main works of this paper are as follows:A set of forest fire smoke dataset was collected and organized through UAV photography and the Internet,which covering multiple dimensions of factors such as quantity,smoke scenes,and disturbance factors;the Label Img tool was used to label the smoke lee in the images with the label smoke and make it into a dataset in VOC2007 format;the smoke data were preprocessed using data enhancements such as flip,brightness,rotation,contrast adjustment,brightness adjustment,and adding pretzel noise to prevent over-fitting of the model.The YOLOv7,SSD and Faster RCNN models,which were pre-trained on Image Net,were migrated to the forest fire smoke scene and their feasibility was investigated.The experimental results show that the training of the models by migration learning can improve the convergence speed and recognition precision,and the migrated smoke recognition model based on YOLOv7 performs best on the smoke dataset with an recognition m AP of 88.01%,which proves the effectiveness and superiority of migration learning.An improved YOLOv7 model was proposed for the problem of poor recognition of thin smoke and small smoke at the early stage of fire by the YOLOv7 model.By adding EPSANet and SANet attention mechanisms to the backbone feature network,the network compensates for the lack of multiscale spatial feature extraction capability;the BIFPN network was used to replace the original PANet network to improve the fusion capability of multidimensional features;the smoke dataset was reclustered based on the K-means++ algorithm to obtain a priori frames applicable to forest fire smoke;the decoupling head was used to decoupling head to decouple the prediction frame classification from the localization task and optimize the loss function to improve the model recognition precision and model convergence speed.The experimental results show that the improved YOLOv7-w model achieves m AP of 93.02% in smoke recognition,which is 5.01% better than the YOLOv7 model.To address the problems of high number of network parameters and slow model inference speed of YOLOv7-w model,two parts of YOLOv7-w model were improved.In the first part,the light-weighted Ghost convolution and DSConv convolution were introduced to reconstruct the backbone network respectively,which significantly reduced the number of model parameters.In the second part,the DSConv-YOLOv7-w network model with better recognition effect was selected,and the pruning algorithm was used to eliminate the redundant channels of this network model,and the pruned model was fine-tuned to improve the smoke recognition effect.Compared with the YOLOv7-w model,the improved YOLOv7-wl model reduced the Params by 4.27 M to 4.3M,increased the recognition speed by 19 f/s to 56 f/s.The experimental result shows that the improved model can efficiently identify the smoke area,reduce the computing requirements of hardware equipment,and have important application value for forest fire monitoring and early warning.