Research Of Fire And Smoke Detection Algorithm Based On Deep Learning

Fire can cause serious damage to public facilities,endanger human life and safety,and cause immeasurable economic and ecological losses.Therefore,it is very important to study fire smoke detection based on videos and images for warnings of fire.In daily life,the scene of fire is complex and changeable.What’s more,fire and smoke have no fixed shape and size,which make it very difficult to implement a fire smoke detection algorithm with high precision.Traditional fire smoke detection algorithms have poor generalization ability,so it is very difficult to achieve high precision and low false detection.In recent years,there are a large of number of fire smoke detection algorithms which using deep learning.Although the precision and efficiency of fire smoke detection algorithms based on deep learning have surpassed the traditional ones,most of them still have the problem of high false detection,making it hard to be applied directly in real life.Therefore,how to implement a fire smoke detection algorithm with high precision and high generalization ability has become an urgent task.Regarding the issues above,the fire smoke detection algorithm will be studied and explored from the perspectives of incremental learning,few-shot learning,multi-scale augmentation and data augmentation.This paper creates a fire smoke dataset(FSD)with real samples and object bounding boxes from news reports.The dataset is rich in scenes,including fire or smoke of different sizes and camera angels.Faster RCNN detection network is used as the base framework.The fire smoke incremental learning detection network(ILOD)is designed on the basis of knowledge distillation.The global attentional feature map is added to the feature distillation to enhance the guiding effect of fire smoke learning.Some meta-learning layers is designed in the Ro I Head Network to accelerate the incremental model convergence.On this basis,a difficult sampling method is proposed to further improve the detection effect of the model.From the experiment results on the fire smoke dataset,the incremental learning fire smoke algorithm proposed in this paper not only get a false detection rate of only 0.63% in complex scenes,but also maintains an accuracy of 80.02 in fire smoke detection.In order to further improve the generalization ability of fire smoke detection and reduce more false detections,a few-shot learning network of fire smoke detection which called Fs Det is proposed in this paper.Firstly,a meta-learning network is used to extract the Ro I regional proposals.And a lightweight reweighting network to extract class-specific reweighting vectors.Then the Ro I regional proposals and reweighting vectors are multiplied according to feature channels.Finally,reweighting Ro I regional proposals will be input the prediction module to predict classes and bounding boxes.In addition,a classifier which based on cosine function is designed to increase inter-class difference and reduce intra-class difference.After the incremental learning network is appropriately adjusted,a few-shot fire smoke detection network based on incremental learning(IL-Fs Det)is proposed,which can further reduce the false detection of fire and smoke while ensuring low-cost labeling work.In view of the variable size of fire smoke objects,a novel multi-scale training strategy is designed in this paper.On a specific training scale,only objects of specific size are allowed to compute loss and update parameters.A three-branches weight sharing selective kernel network(TWS-SKNet)is designed to increase the receptive field for objects of different scales.In additional,dilated convolution and weight sharing are used to reduce the amount of parameters.For small fire smoke objects,the detection performance of small objects improved by copying and pasting data augmentation.Finally,multiple sets of comparative experiments are used to explore data augmentation methods which are suitable for fire smoke detection,further improving the generalization ability of the fire smoke detection model.