Research On Smoke And Fire Detection Method Based On Deep Learning

With the development of the social economy and the widespread adoption of various scientific and technological equipment,urban environments are becoming increasingly complex.As a result,detecting smoke and fire is becoming more difficult.Accurate and rapid detection of smoke and fire can provide early warning during the early stages of a fire,enabling people to take precautions in advance and minimize the impact of fire on the living and working environment.Traditional methods based on sensors and machine learning have limitations when dealing with complex environments and large-scale use.With the increasing maturity of deep learning technology,smoke and fire detection methods based on deep learning are receiving more attention.However,challenges remain,such as low accuracy of detection,difficulty detecting small targets,and limited generalization ability.Regarding the problems of smoke and fire detection algorithms,this article conducts research on smoke and fire detection methods based on deep learning.The main research work and innovative aspects of this article are reflected in the following aspects.(1)A smoke and fire image dataset was built.By collecting a large number of smoke and fire images from multiple publicly available datasets,as well as adding screenshots of fire scene videos to complete the initial dataset.However,at this time,the quality of the dataset was not high enough,with duplicated content,low image resolution,and empty labels.To ensure efficient training and model generalization,a structural similarity algorithm was used to remove duplicates from the dataset.In the end,without data augmentation,the dataset consisted of 9981 smoke and fire images,with small target data as the main focus,covering different scenes,angles,lighting,and other factors.Additionally,these images were meticulously annotated and categorized to make them more usable and operable.Building our own dataset helped reduce the limitations of the dataset on the model as much as possible.(2)An improved YOLOv5 s backbone network based on the SimAM module is proposed.The YOLOv5 s backbone network is augmented with a 3-D attention mechanism module called SimAM.By capturing the interaction between different features,this improves the feature representation ability of the network and ultimately the performance of the model.This attention mechanism can weight the features in the network without adding additional parameters,enhancing the representation ability of key features and improving the performance of the network.(3)Propose an improved YOLOv5 s feature fusion network based on BiFPN for detecting smoke and fire.The three-scale detection in YOLOv5 s is changed to four-scale detection to enhance the detection ability of small smoke and fire targets,but interference is introduced into the feature map.To optimize the feature fusion process,the weighted bidirectional feature pyramid network(BiFPN)structure is adopted,which further enhances the feature fusion ability through cross-level connections and trainable weight mechanisms.The model’s detection accuracy and feature fusion ability for small targets are improved,thus further improving the algorithm’s accuracy.(4)Propose a genetic algorithm-based approach for optimizing hyperparameter combinations in the improved YOLOv5 s.The approach is applied to optimize the hyperparameters of the modified network on the self-built dataset,including learning rate,loss coefficients,etc.By using genetic algorithm,the optimal hyperparameter configuration can be found more efficiently,leading to improved performance and generalization ability of the model.By adding the SimAM module and optimizing the feature fusion network with BiFPN,the feature extraction and fusion capabilities of the network have been effectively improved.The optimal hyperparameters found by the genetic algorithm also improved the detection performance of the model.Experimental results show that the improved smoke and fire detection algorithm based on YOLOv5 s proposed in this study achieved an average detection accuracy improvement of 7.2% compared to the original algorithm under almost the same detection time.The accuracy of detecting small targets has been improved,and the anti-interference and generalization abilities of the model have been enhanced,greatly reducing missed detections and false alarms.These results demonstrate the significant advantages of the proposed approach in smoke and fire detection,which can accurately detect smoke and fire information at the fire scene,providing important support for the rapid response and effective disposal of fire accidents.