Research On Fire And Smoke Recognition Method Based On Deep Learning And Multispectral Fusion

Fires have enormous destructive power,often causing extensive casualties and property damage.Thus,timely and effective smoke and flame detection can issue an alarm in the early stages of a fire,reducing the probability of a fire occurring,and greatly reducing the property damage and casualties caused by a fire.Traditional fire detection methods,such as smoke and temperature sensors,lack effectiveness and real-time capabilities in complex environments such as outdoors,forests,and large spaces,making it difficult to meet the needs of large-scale,high-precision fire detection.Moreover,traditional visual fire detection systems are prone to misidentification or omission in complex environments.Therefore,this paper proposes a multi-modal fire detection algorithm based on 3D convolutional neural networks.This algorithm combines visible light and near-infrared images,and uses a 3D VGG model for feature extraction and classification to achieve accurate identification of fires.Furthermore,to verify the feasibility of the algorithm,this paper developed a multi-modal fire detection system that uses deep learning technology and multi-modal image fusion technology,which can quickly and accurately monitor and identify fire situations.Firstly,due to the lack of multi-modal fire datasets,this paper conducts fire simulation experiments based on smoke and flame samples,and interference samples selected from fire simulation experiment in accordance with national standards.The paper collects multi-modal video data of smoke and flame in near-infrared and visible light when a fire occurs,and uses a manual calibration method for image registration.After calibration and processing,this paper produces 5377 video fire samples.To fully train and evaluate the model,3450 samples are used for training,863 samples are used for verification,and 1064 samples are used for testing in this dataset.Secondly,this paper proposes a fire recognition model based on deep learning algorithm,called Att-3DVGG.Studies have shown that the dynamic features of fires play a critical role in fire recognition in complex backgrounds.Therefore,this paper uses 3D convolution to simultaneously extract dynamic features between video frames and static features within frames.Additionally,because specific information in visible light and near-infrared images can complement each other,multi-modal image fusion is used to improve the recognition rate of fires.Experimental results show that the accuracy of this model in identifying fires reaches 99.90%,which is significantly higher than other video-based fire recognition algorithms(such as 6M3 DC,Des Net3 D,etc.),and has high recognition accuracy and robustness.Lastly,this paper constructs a complete multi-modal fire detection system that runs on Windows.The system receives real-time video streams of near-infrared and visible light as input and can register and fuse near-infrared and visible light video streams in real-time to achieve real-time fire detection and alarm.Experimental results show that this system not only efficiently fuses video data but also effectively improves the accuracy and stability of fire recognition.This system provides a powerful tool for fire monitoring and helps prevent fires,ensuring the safety of human life and property.Figure [25] table [6] reference [60]...