| Most of the chemical raw materials stored in chemical plants are flammable and explosive.Once the warehouse storing these chemical raw materials is burned,it will cause explosion and serious consequences.Therefore,chemical plants must take fire prevention and disaster prevention measures to do a good job of prevention.Using flame detection technology to avoid risk is a good choice.Fire detection has been widely used in production safety and monitoring and other related fields.Detection in the early stage of fire can minimize the damage of fire to life and property,and reduce the damage to ecological environment,social economy and other aspects.With the development of digital image processing and pattern recognition technology,Video Fire Detection(VFD)has become an important means of Fire Detection.People can collect the continuous motion information and image information of the detected object in real time through the camera,so as to obtain more abundant flame characteristics,and directly detect the flame through digital image processing and pattern recognition methods.Compared with traditional fire detection methods,fire detection based on computer vision features has better accuracy and environmental stability.However,due to the complexity of application scenarios and some technical difficulties,the performance of video flame detection is not high enough,so it is worthy of further research and application.In this paper,a new video flame detection method is proposed based on deep learning model.(1)A YOLOv2 video flame detection method based on multi-level feature fusion is proposed.In order to solve the problem of target loss caused by the small subsampling resolution,a deconvolution module was introduced,and the features with strong semantic information in the deep layer were integrated with those with strong detailed information in the shallow layer,so as to effectively improve the detection rate of the flame.(2)A light weight YOLOV4 video flame detection method is proposed.In view of the problem that the Model has a large number of feature graph scale changes,leading to the slow Model operation speed,the PANET structure in the Model was removed,and a Fire Model module was proposed to strengthen the connection between the upper and lower convolutional layers of the Model,to compensate for the inadequate semantic information fusion after the removal of the side connection with the PANET,and to improve the flame detection rate.The model in this paper is trained under the same conditions with other mainstream deep learning models,and the same test set is detected.The performance of the model is judged by the detection results of different models for flame video and non-flame video.The experimental results show that the model presented in this paper has high detection accuracy and strong practicability. |
