Research On Fire Image Extraction And Recognition Based On Machine Learning

As a kind of frequent disaster accident,fire poses a great threat to the safety of people's lives and properties.Therefore,it's of great significance to the monitoring and the early warning of the early fire.The traditional fire detection technology is by means of the smoke,light,heat and other sensors,which has many disadvantages,including the limited detection range,long response time and low accuracy.With the popularity of the surveillance cameras and the rapid development of the image processing and pattern recognition technology,the fire detection technology based on video images has come into being.Compared with the traditional fire detectors,this technology has obvious advantages in many aspects,such as the alarm speed,the visualization,the wide coverage and so on.Throughout the existing Image-type fire detection technologies,mostly are based on the flame image features recognition.Firstly,this paper has optimized and improved the calculation method of some flame image features.Besides,due to the influence of noise,illumination,the equipment and other conditions,the existing algorithms has a low accuracy of the flame foreground extraction,leading to the poor robustness of the whole system.Basing on this background,this paper has proposed a new flame image foreground extraction algorithm.It firstly uses the supervised learning method to determine the suspected foreground region,then segments the image blocks corresponding to the suspected area with the improved K-means clustering algorithm,finally obtains the relatively accurate flame foreground images.Without a public authoritative fire video dataset for support,this paper has conducted a large number of simulation experiments and recorded the videos,also collected some other peoples' publicly available fire videos from the Internet,self actively organized related data and constructed a fire video dataset.Based on the established dataset,this paper has implemented the foreground extraction algorithm mentioned above.The experimental results show that the algorithm can accurately extract the foreground of flame images and keep rich flame image information with guaranteeing the speed of the algorithm,and the robustness is stronger.Finally,a convolutional neural network model for flame image recognition is constructed,and a video image fire detection system based on deep learning is designed combining with the suspected flame region extraction method.And the training and testing are carried out by using the self-built dataset.The experimental results show that this system can reach a high detection rate of more than 85% with a stronger robustness and a higher practicability.