Research On Fire Detection Algorithms Based On Multimodal Images

The occurrence of fire directly endangers the life and property safety of the general public.Traditional fire detection uses a variety of sensors such as light,smoke,and temperature.However,this method is limited by many factors such as detection distance and equipment installation location,and has problems of slow detection speed and missed detection and false detection.With the development of computer vision technology,the fire detection technology based on video image realizes the detection of flame and smoke through the feature information in the video and the image,which has the advantages of fast response speed,wide coverage,timely information storage,and rich alarm forms.However,this method has limitations.On the one hand,the public pyrotechnic datasets are insufficient and difficult to obtain,which makes the model’s expression and generalization capabilities insufficient to meet practical application requirements.At the same time,the scale span of flame and smoke in fire images is large,and it is difficult for existing algorithms to extract multi-scale feature information.On the other hand,it is impossible to completely capture the appearance,luminescence,heat generation and other characteristics of fire only by using single-modal images for detection.For example,visible light images are easily disturbed by external light,resulting in serious aliasing of fireworks and backgrounds;near-infrared images can effectively highlight the active light-emitting characteristics of flames,but are easily interfered by other active light-emitting light sources.In response to the above problems,this paper uses the multi-scale feature expression in deep learning to extract more discriminative features in flames and smoke at different scales,and uses a multi-modal image fusion mechanism to make up for the shortcomings of single-modal image detection.The methods and innovations proposed in this paper are as follows:1)A fire detection algorithm based on multi-scale features is proposed.For the problem of large-scale span,a multi-scale feature extraction module and a multi-scale feature fusion module are proposed.By using multi-branch and asymmetric extraction mechanism in the feature extraction process in lightweight detection network to extract multi-level feature information and building multi-scale dilated convolution layers to fuse smoke and flame multi-scale information.Finally,the reinforcement learning of multi-scale features and the increase of the network receptive field are realized.In addition,in view of the problems of insufficient samples and single scene in the existing fire data set,this paper builds a fire image data set DFS for model training and testing.The proposed method achieves 98.5% detection accuracy on the DFS dataset,with only9.2M model parameters,achieving competitive detection performance.2)Design a fire detection algorithm based on multi-modal image fusion,which can effectively fuse the apparent characteristics of flames in visible light images and the active luminescence characteristics in near-infrared images.Considering that the fire information has different characteristics in the visible light mode and the nearinfrared mode,but in the shallow feature,the fire information in different modes is still related.Therefore,based on the fire detection algorithm based on multi-scale features,this paper uses parameter migration to preserve the shallow network weight parameters in the visible light modal detection model.With the help of near-infrared modal image data,the weight parameters of the deep network are learned,and the deep feature information is obtained.Finally,a decision fusion algorithm is designed,which uses the fire detection network of different modalities to obtain target information in different modalities,and uses the decision fusion algorithm to fuse the location information and confidence results of the two modalities.The detection accuracy of the algorithm on the self-built DVN dataset,Corsican Fire Database and RGB-NIR Scene Database is 98.8%,which is 0.8%-13.9% higher than other algorithms.Experiments show that the algorithm can make good use of the characteristics of multimodal images and effectively improve the detection accuracy in different environments.