Dynamic Smoke Detection Based On Convolutional Neural Network

The occurrence of wild fires will cause huge economic losses and even endanger human life every year.Due to the large area of the wild forest and strong environmental interference,the sensor-related fire detection technology has the disadvantages of high cost,limited detection range and easy aging,so it is not suitable for the scene of wild fire detection.However,with the development of video surveillance technology and the popularization of related technologies,research on field fire detection has gradually begun to develop in the direction of video recognition and detection,which makes dynamic smoke detection technology of important research significance.In order to expand the viewing angle and scope of field monitoring as much as possibles,the monitoring must obtain long-range images.When a fire occurs,the first feature that appears in the monitoring range is smoke.Therefore,the research of dynamic smoke detection methods has a very positive and important role in fire warning and rescue response.Since there is no unified data set for video smoke detection,this study first collects and sorts out the smoke data set.In this process,the smoke image is filtered,clipped,normalized and other processing to form a data set.In order to avoid the blocky distortion of the smoke image during the zooming process or the unsatisfactory dynamic smoke recognition effect caused by "mosaic",the nearest neighbor interpolation algorithm,bilinear interpolation algorithm,bicubic interpolation algorithm and other methods are adopted to suppress block distortion.Through experimental tests,it is concluded that the bilinear interpolation algorithm has more advantages in smoke image recognition than the other two algorithms.Therefore,the bilinear interpolation algorithm can be used as a scaling algorithm for dynamic smoke image detection,which can reduce the missed detection rate while ensuring the real-time detection effect.Secondly,in the study of the algorithm for extracting the suspected smoke area,it is found that the traditional method of using the Gaussian mixture model or the inter-frame difference method to extract the dynamic smoke area often requires manual feature selection to determine the suspected smoke area.Studies have shown that when the smoke continues to appear,the traditional Gaussian mixture model will determine the smoke area as a background area,resulting in a missed detection phenomenon;while the motion area extracted by the frame difference method will have holes.Therefore,based on the traditional optical flow method,this article uses the method of optical flow estimation on the corner points of the video smoke image to extract the moving area,and at the same time,by identifying the main moving direction of the area and extracting the HSV color space characteristics of the smoke,the suspected smoke is determined area.The improved extraction method of suspected smoke regions in this paper can improve the accuracy of dynamic smoke detection.Finally,this paper compares and tests the recognition model,chooses the convolutional neural network as the recognition model of smoke detection,and realizes the real-time detection of dynamic smoke.