Research On Intelligent Flame Recognition Based On Transfer Learning

With the development of society and the accelerated pace of people’s lives,cameras have been placed in large numbers in various places,and special attention has been paid to fire safety issues.It is not realistic for a large number of cameras to be manually monitored.Therefore,the automatic detection and identification of fires is particularly important.It is very meaningful and valuable to automatically identify and detect fires based on video and images.Traditional image processing method has made considerable progress for the processing and discrimination of such videos and images,but there are still some deficiencies.In recent years,machine learning,especially the emerging technologies brought by neural networks,have brought great technological innovations to this kind of video-based automatic detection.So,in this paper,we proposed two methods based on transfer learning to identify the flame in images.First we proposed a method used fine-tuning after transfer learning,and compared to no fine-tuning method,our method has greatly improved the recognition accuracy.The other method we proposed is that we used transfer learning models to extract the features in image,fused features and then input to machine learning classifier to identify flame.In the proposed procedure,two ensemble learning methods,boosting and stacking,are used for experiment,the experimental results show that the AUC areas is close to 1,and the accuracy reached up to 96.5%.So,our proposed method can achieve good results in flame identification under the complex background environment of fire video.And then,we proposed a method that based on YOLOV3 algorithm,we used k-means algorithm to change the anchors,and used some training trick to make the algorithm more robust,such as change the input shape of the image,change the number of categories first and then set the categories to 1,experiments shows that can reduce the rate of missed detection,and compared to the image processing method,this method can get better results than these methods.Our proposed method used to recognize video,frames rate can exceeds 50 frames per second.So our method can also meet the requirements for real-time detection and analysis.