Research On Fire Image Recognition Method Based On Transfer Learning

In social life,fire has seriously threatened public safety and damaged human property and life safety.How to detect and warn fire as soon as possible is always the focus of people’s research.Traditional fire image recognition based on image processing was through feature analysis and extraction of flame or smoke images to identify the fire.The feature extraction algorithm designed by man was difficult to accurately extract features from multi-background complex images,and the algorithm also had problems such as low recognition accuracy and poor real-time performance.The current fire image recognition based on deep learning is to use deep neural network to learn samples from fire pyrotechnic data set and train the network that can identify flame or smoke to achieve the purpose of fire recognition.However,this method also has some problems such as less data set and time-consuming training.In order to solve these problems,a fire image recognition method based on transfer learning is proposed in this paper.Transfer learning is applied to the field of fire recognition,and transfer method of shared parameters is used for fire recognition.The research work of this paper is as follows:（1）For the algorithms of fire image classification and fire target detection,two deep frame environments,Tensorflow and Torch,had been deployed and the corresponding data sets had been established;（2）In the fire image classification and recognition,data enhancement methods such as clipping,flipping and color transformation was used to preprocess corresponding data sets to prevent the phenomenon of over-fitting;The pyrotechnic data set required for target detection algorithm training was labeled and made into YOLO format data set;（3）With Xception,Efficient Net,VGG16 and other classical deep learning networks trained in Image Net for transfer learning,Efficient Net network had the best performance in flame data set,optimized and added CBAM attention mechanism.The flame recognition rate had reached 98.3%;（4）Target detection technology was applied to the field of fire,a fire recognition method based on YOLO had been proposed,and its structure was adjusted,Neck part was replaced by BIFPN structure of the original PANET,in order to integrate more features without increasing too much cost.The experimental results showed that the model can accurately and quickly recognize the fire image and mark the location information of the fire;（5）In order to make the algorithm application go out of the laboratory and be applied in the actual scene,the trained model YOLO-v5 file was deployed on the embedded platform NVIDIA Jetson Nano,the real-time picture was read through the external camera,and the parameter accuracy was reduced to the semi-precision floating point type by using Tensor RT to accelerate the network model inference.In the case of less loss of accuracy,the operation speed was greatly improved,and the test system was applied in fire identification.Finally,the average test frame rate of the system was 13 FPS in the pyrotechnic data set,which can complete the real-time processing of the monitoring video stream information and accurately identify the fire.