Research On Flame Detection Method Based On Deep Neural Network

Fire is a disaster caused by uncontrolled combustion in time or space.Once it occurs,it can bring great suffering and loss to human beings,so timely and accurate detection of fire can effectively protect human life and property.At present,the detection of fire in society mainly uses the characteristics of flame,smoke and temperature in the process of fire,among which computer vision processing of flame images can effectively use the characteristics of fast light propagation,which is an effective method to detect fire flames.Facing the flame target image with complex background and variable scale,the traditional computer vision-based flame detection method requires complex operations such as manual extraction of features,which is slow and poor in detection accuracy,and the robustness of the detection method is insufficient to meet modern needs.The deep neural network-based target detection method can use its own powerful abstract feature learning ability to detect and locate flame targets from data autonomously,thus gradually becoming a new trend in flame detection methods.Various deep neural network-based flame target detection methods have been designed for the modern application of flame detection methods.The main work of this paper is as follows:(1)The basic principles and functional structures of deep neural networks are introduced in detail.Then typical target detection models are reviewed,and a flame detection dataset that can meet the training of deep neural networks is established,and the evaluation metrics involved in target detection are elaborated.(2)To address the problem that the existing flame detection models have a complex structure and a large number of parameters,which are difficult to deploy in lightweight devices.Firstly,the inverted residual module containing Channel Shuffle is introduced to extract the flame target features;secondly,the combination of depth-separable convolution and point-by-point convolution is used to replace the ordinary convolution responsible for downsampling in the original network to further reduce the complexity of the network structure;finally,the C3-s module in the Neck network structure is improved and the Ghost module is introduced to replace the ordinary 3*3 convolution to reduce the computational effort required for fused feature map feature extraction.It is experimentally verified that the improved model ensures the same detection accuracy as YOLOv5 l while the volume of the network model is reduced to 34.8% of the original one,which is an effective lightweight flame detection method.(3)To address the problem of low accuracy of existing target detection models for flame in industrial deployment,a high accuracy flame detection model based on multiattention mechanism and spatial feature fusion is proposed.By introducing coordinate attention mechanism in the backbone network,the relationship between flame position information and channel information is obtained in an efficient way to fuse flame position feature information and channel feature information.The Swin Transformer Block is introduced in the Neck network to expand the perceptual field of the network model and improve the flame feature extraction capability.The adaptive spatial feature fusion module is introduced into Head network to enhance the multi-scale flame spatial feature fusion and improve the detection accuracy.The improved network model improves the average accuracy of the model by 4.1% compared with the original YOLOv5 l.Compared with other 11 existing detection models,the best average accuracy of 66.8% was achieved,which is a high accuracy flame detection model.