Research On Algorithm Of Deep Learning In Video Flame Detection

The occurrence of fire will have a disastrous impact on the safety of people's lives and property and social stability.Especially in open scenes such as forests,substations,and large warehouses,temperature and smoke are easily weakened in the air flow process.The use of traditional fire sensors to detect fires has low accuracy and even missed detection.Therefore,in recent years,based on video image processing Fire detection is an important research direction in the field of fire monitoring.The research object of this paper is video flame data,and two video flame detection models are designed using image enhancement and deep learning methods.The two models in this paper can not only reduce the manual extraction of features,but also deeply dig into the dynamic and static characteristics of the flame in the video.The recognition rate and real-time performance of the video flame detection are improved,and the false alarm and false alarm rates are reduced.(1)In order to solve the problem of low video image quality and the traditional histogram enhancement will bring excess noise,an improved CLAHE image enhancement algorithm is proposed based on the limited contrast histogram(CLAHE)algorithm,which is based on the degree of uniformity of gray levels.Adapt the setting of the cutting point position.The image processed by the improved CLAHE algorithm using image quality evaluation proved that the noise in the homogeneous region is not obvious,and the information is richer and more readable.(2)In this paper,an improved dual stream convolution(ITCN)video flame detection model is constructed.The model extracts the dynamic characteristics of the flame and the static characteristics of the space through the time flow convolution and space flow convolution networks,respectively.The large convolution kernel of the model is replaced by a small convolution kernel to reduce the amount of model parameters and calculation,and the O-inception module is added to the model to enhance the model discrimination ability.The experimental results show that the model uses concat fusion to obtain the best network feature fusion effect,and the model can achieve 95.15% accuracy while meeting the real-time performance of the system.The model's generalization ability test on public data sets also has good performance.(3)In order to dig deep into the long-term motion characteristics of the flame time dimension and enhance the reliability of the flame detection model,this paper builds a video flame detection model based on an improved dual-stream convolutional network and long-short-term memory network(ITCN-LSTM).Based on the improved dualstream convolutional network,the model adds two layers of LSTM networks,each with 128 hidden units.Therefore,the time series data processing capability of LSTM can be fully utilized.The experimental results show that the comprehensive performance of the ITCN-LSTM video flame detection model is better than the ITCN model and other algorithms in performance comparison experiments,fault tolerance test and generalization test experiments,which verifies that the ITCN-LSTM model not only has high recognition rate and low False positive rate,low false negative rate,and also has excellent fault tolerance.The reliability of the model has been improved,which further demonstrates that the ITCN-LSTM model can dig deeper into the dependencies between video flame frames.