Fire Smoke Warning Algorithm Based On Machine Learning And Computer Vision

Smoke identification is of great significance for the early detection of fire and the reduction of fire hazards.At present,there are many false alarms in smoke detection based on computer vision,and the model will be too large when using deep network.In order to solve these problems,in this paper,we put forward the incremental judgment algorithm of smoke picture block,and builds a lightweight deep neural network model to identify smoke.In this paper,firstly,through the network way to collect smoke pictures,relying on the fire smoke video collected by Tianjin Fire Control Institute,we select and intercept the smoke pictures,then use the data enhancement method to expand the number of pictures,and establish the smoke data set,including 12000 smoke pictures and non-smoke pictures respectively and several smoke videos.In order to reduce the interference of similar smoke targets and improve the accuracy of the system,we proposes an incremental decision algorithm based on smoke image block.Firstly,the smoke is extracted and divided into blocks by background subtraction method to eliminate background environment interference and reduce calculation and recognition difficulty;secondly,RGB color model is built to eliminate non smoke interference and extract suspected smoke block according to the gray prominent characteristics of smoke color;considering the rapid increase of smoke in the early stage of fire,the suspected smoke picture blocks are identified by whether the number of smoke picture blocks is increasing.By analyzing and extracting reasonable experimental parameters,combined with the block method,the early warning timeliness is ensured,and the false alarm caused by cloud,fog,haze and other smoke analogues is effectively reduced.When choosing the recognition model,AlexNe,VGGNe,Goog Le Ne,Res Net and MobileNet model are constructed and studied respectively.To optimize the MobileNet,we use the transfer learning method and the Focal Loss function,to reduce the contribution of easy to classify samples to the loss function,and improve the accuracy to 0.989.When reducing the model volume and parameter amount to 1/5and 1/14 of the AlexNet,the accuracy reachs or even exceeds the accuracy of the conventional neural networks,and reduces the calculation time for smoke The fog recognition model lays the foundation for deployment to devices with small memory.In this paper,we combine the incremental decision algorithm of smoke image block and the improved mobilenet lightweight model to identify the fire smoke,which effectively reduces the volume and calculation of the model,reduces the non smoke interference,improves the accuracy and sensitivity of the smoke identification,w provides a theoretical basis for the application of the smoke identification system.