Research On Convolutional Neural Network Image Detection Algorithm For Electrical Fire Prevention

With the continuous improvement of the electrification of modern society,the probability of occurrence of electrical fires has increased year by year,and the damage caused by it has become more and more serious.Therefore,various methods for early warning of fire have emerged.Image recognition,as one of the important means to achieve fire warning,has great advantages in real-time and visualization of early warning.However,traditional image recognition methods have low detection accuracy and are difficult to identify small targets.Aiming at these problems,this paper combines the image algorithm based on convolutional neural network to deeply study the small target recognition and detection of fire prevention and control.In view of the shortcomings of current fire identification algorithms and the shortcomings of fire detection algorithms,combined with the excellent performance of convolutional neural networks on visual tasks,a fire image recognition algorithm based on hybrid convolutional neural network and a fire target based on improved YOLOv3 are proposed.The detection algorithm is designed according to the fire identification detection algorithm proposed above,and a fire detection system is designed to meet the actual needs of real-time monitoring of fire hazards.Among them,the main research contents and results of this paper are summarized as follows:(1)For the traditional fire image recognition algorithm,the detection accuracy is low and it is difficult to identify small targets.In this paper,the feature extraction method of convolutional neural network is deeply studied,and combined with the feature fusion method,a hybrid based method is proposed.Fire identification method for convolutional neural networks.The method extracts the feature information of different scales by mixing convolutional neural networks,and improves the recognition accuracy of fire by merging the feature information of different scales.(2)In view of the lack of the current public fire data set,this paper constructs a multi-scenario large-scale fire target detection database.Existing fire detection methods treat fire detection as a two-category problem,enabling classification of fires and non-fires.In this paper,fire detection is formalized into multi-classification and coordinate regression problems based on fire and smoke information.In view of the shortcomings of the current target detection model in small target detection performance,this paper proposes a fire detection method,which improves the problem of insufficient YOLOv3 small target detection capability and performs on the fire target detection database constructed in this paper.Offline training has verified the feasibility of deep learning in detecting small fire targets.(3)Development and application of fire detection systems.Based on the two algorithms mentioned here,a real-time fire detection and early warning system is developed.The system can display the fire detection result of the camera monitoring screen in real time through the GUI interface.The operator can control the remote network camera according to actual needs.Various parameter,including screen focal length,pan/tilt steering,and control of local fire detection switches.