Forest Fire Smoke Recognition Based On Deep Learning

The forest resources are extremely important for human development.Fire has the characteristics of strong sudden and high frequency,which not only threatens the safety of personnel,but also causes huge economic losses.Therefore,fire prevention has become the primary task of protecting natural resources in various countries.The smoke is the early signs of fire,accurate and timely identification of smoke,to reduce resource damage and protect the ecological resources,has an important significance.Fire smoke detection usually arranges visual sensors at detection points,collects pictures or video information in real time,and then judges whether there is a fire manually,which has the disadvantages of high labor cost and high missed rate.In view of the above problems,this paper intends to use deep learning algorithm to train the self-built data set and adjust the deep network model from two perspectives of image recognition and video recognition,in order to explore the feasibility of the application of deep learning technology in forest smoke monitoring scenarios.The main work of this paper is as follows:(1)Data set establishment.The smoke image data set and video data set are established mainly by combining manual acquisition with current open database.(2)Based on the image.In-depth study of the network model structure of in-depth learning.Faster R-CNN and SSD network are used to detect smoke image.The experimental comparison shows that the accuracy of Faster R-CNN network is better than SSD network in the field of smoke recognition.(3)Based on the video.Considering the time series relationship between image frames and frames in forest fire smoke video,a video-based forest fire smoke detection model is proposed.Verify the detection accuracy of different CNN+RNN network structures for forest fire smoke,and determine the optimal model network parameters.Finally,the improved VGG16+GRU network model is better for smoke recognition,and the recognition accuracy reaches 75%.Finally,a visualization platform of forest fire smoke recognition results considering in-depth learning is established.