Research Of Fire Detection Based On Significance Detection And Convolutional Neural Network

With the deployment of industry and the expansion of urban areas,early warning systems play a vital role in providing emergency situations.In production and life,fires can’t be prevented,and they can spread quickly and cause injury and death.Therefore,it is very important to provide early warning for automated fire detection tools.Traditional detectors usually use photometry or ionization to distinguish the presence of flame particles.These methods are suitable for indoor applications,but they are very sensitive to any small fire or smoke and the number of false positives can be high.Another important disadvantage is their limited distance.This means they only alarm when fire or smoke irritates them,and we can’t detect them remotely in real time.With the development of camera technology and the improvement of storage capacity and computing power,video-based fire detection systems have replaced the traditional fire detection methods as research hotspots in the field of fire identification.By using video image-based image processing or computer vision technology,we can identify out-ofcontrol fires at an early stage before a fire becomes a disaster.This paper first uses image processing technology to segment relevant regions,feature extraction and analysis,and feature recognition,and then further learns flame and smoke related features through convolutional neural networks in deep learning methods.Finally,the relevant methods are compared and verified in the public data set and the own data set.We use spatial and temporal features to improve accuracy.Divide the captured video sequence into spatiotemporal blocks.Based on this,the method of combining saliency mapping and color and texture features is used to detect the fire area.We use the HSV color model as the spatial feature and LBP-TOP as the temporal processing of the flame texture.Fire detection experiments on public data sets show the accuracy and robustness of the algorithm.Due to the limitations of traditional image processing in artificial description of related features and extraction,deep learning has huge advantages over traditional image processing.Therefore,this paper proposes a dual-channel convolutional neural network that can detect both flame and smoke.First,extract the relevant image frames in the video sequence and use the convolutional neural network to extract the relevant features layer by layer to avoid the limitations of the artificial design feature extraction method.According to the characteristics of flames and smoke under normal circumstances,classification features are extracted from different convolutional layers.Through the phased training of the detection module and the classification module,an end-to-end convolutional neural network is finally realized.