Research On Image-based Fire Detection Algorithm Based On Lightweight Convolutional Neural Network

Fire is a very destructive disaster in daily life,timely detection and alarm is of great significance.Compared with traditional sensor-based fire detection methods,vision-based fire detection methods have many advantages,such as high accuracy,fast response speed,wide application range and abundant alarm information.In the early stage,image-based fire detection methods were based on traditional features such as color,shape,and texture,which were difficult to adapt to fire detection in complex environments.With the rapid development of deep learning technology,fire detection methods based on convolutional neural network and object detection have gradually become a research hotspot.However,existing fire detection methods based on deep learning failed to achieve a good trade-off among detection accuracy,false alarm suppression and model complexity.Therefore,this thesis proposes a fire detection method based on lightweight convolutional neural network.The lightweight fire detection network takes FCOS as the basic network,reduces the number of channels and introduces Ghost Net as the backbone,so that the network becomes lightweight.In the backbone,dynamic convolution is introduced to improve the feature extraction ability of the changeable flame without increasing the width and depth of the neural network.The spatial attention module is added to optimize the expression of network spatial features,and H-Swish activation function is used to optimize the learning ability of network.For loss function,the color weight is used to improve the classification loss function and increase the focus on the flame color area during the training phase.Center-ness,as the weight of regression loss,is used to improve the contribution of central region features.In addition,the motion foreground detection is introduced before the fire detection network to form a complete fire detection algorithm.In this thesis,a fire detection dataset with rich scenes and standard labeling is established,and a series of ablation experiments and comparison experiments are conducted on self-built dataset and public dataset.The average precision of the lightweight fire detection network in the self-built fire dataset is 92.0%,the amount of parameter is 4.58 M,and the floating point operations is 31.45 G.The complete fire detection algorithm has an average true positive rate of 99.23% for fire videos,and an average true negative rate of 99.21% for interference videos in the public dataset.In the real monitoring scenario for a long time,the algorithm has no false alarm and the response speed is fast.Experimental results show that the proposed algorithm has good detection results for various fire scenes,and has advantages in detection accuracy,false alarm suppression and model complexity,which has high application value.