Research On Mine Fire Flame Recognition Method Based On Video Image

Mine fire has the characteristics of strong destructiveness and catastrophicity.Once it occurs,it may have serious consequences.Effective monitoring of mine fires is an important prerequisite for ensuring safe production in mines.Traditional monitoring methods such as gas,temperature or smoke are costly and inefficient,and the scope of use has certain limitations.With the intelligent construction of mines,the video images obtained by robots or surveillance cameras can realize fire identification,which has the advantages of wide monitoring area,rich and intuitive information and strong universality.In this paper,digital image processing and deep learning technology are used to study mine fire recognition based on video images.The main research contents are as follows :(1)Mine video image feature analysis and image preprocessing.By analyzing the characteristics of mine video image,it is found that the mine video image has the characteristics of uneven distribution of brightness,low clarity and low color richness.The image is preprocessed by histogram equalization and median filtering.The defogging algorithm of mine image is studied.It is found that the ACE defogging algorithm has a good defogging effect,which greatly improves the definition of dust and fog image.However,there are still foggy areas similar to fire smoke in the image,which provides the basis for using flame image to identify mine fire.(2)Research on mine fire flame recognition based on multi-feature.The research on multi-feature extraction of flame is carried out.The video of fire-like target is collected in a large metal mine in China.The color features,texture features and shape features of flame and fire-like objects are extracted and compared.It is found that some interferences can be excluded through these features.A multi-feature mine fire flame recognition method is proposed.HOG features are extracted from suspected areas that meet the flame characteristics,and SVM is used for flame recognition.(3)Improved YOLOv5 s mine fire flame recognition research.Aiming at the defects in the original network of YOLOv5 s mine flame detection,CBAM attention mechanism is added to the backbone network to improve the learning ability of the network to image features.The CIOU loss function is used to solve the problem that the original loss function has the limitation of aspect ratio in the regression loss.The Sim OTA dynamic sample allocation strategy is adopted to improve the problem of poor training results caused by the original model ’s forced increase in the number of positive samples.The effectiveness of the improved method is verified by ablation test and comparison test with other flame detection networks.(4)Mine fire flame recognition simulation.The performance of the mine fire flame recognition method proposed in this paper is tested in practical applications.The mine environment is simulated,and the flame video images under different conditions are identified.The results show that the two mine fire flame recognition methods can effectively identify the flame,and the improved YOLOv5 s mine flame recognition method can be used for real-time recognition.It shows that the mine fire flame recognition method proposed in this paper has certain practical feasibility.This paper studies the flame recognition method of mine fire based on video image,which has certain reference significance for the research and application development of mine fire monitoring.