Research On Smoke And Fire Image Classification Algorithm Based On BAN

Image classification is one of the important tasks in the field of computer vision.It refers to the image processing method that achieves the classification of different categories by acquiring the salient features of the image.In recent years,with the development of the Internet and communication technology,the number of pictures has increased geometrically.Due to the disorder and diversity of pictures,it is difficult for traditional image classification methods to accurately and completely obtain effective information.Therefore,using convolutional neural networks and deep learning to accurately classify images according to users’ needs is of great significance for both individuals and enterprises.Frequent fires seriously affect the development of our country’s economy and often cause serious harm and cause serious loss of life and property to the people.In recent years,image classification technologies such as deep learning and machine learning have been extensively researched and applied.They can well assist smart devices such as video surveillance to monitor fire areas more easily.However,the manual recognition and classification of smoke and flame images requires a lot of energy and time,and there may be problems such as high labor intensity and strong subjectivity.With the help of deep learning technology,the classification accuracy of smoke and flame images can be greatly improved.In order to improve the efficiency of image classification of smoke and flames in fire scenes,a Classification Algorithm of Smoke and Fire Image Based on BAN(Bilinear Pooling with Attention Network)is proposed.The algorithm is divided into three steps: First,the model uses the color features of the fire or smoke images to pre-classify the input images.If the preclassification fails,it is judged that no fire has occurred,and subsequent classification is not necessary.If it passes,the Res Net-50 is used as a basic network to extract features of fire pictures or smoke pictures.Then,two attention mechanisms,namely Channel Attention and Spatial Attention,are combined to extract key content information and location information from the picture.Finally,use the bilinear fusion module to merge them to achieve classification.Experimental results show that the model can more effectively improve the classification rate of smoke images and flame images,and can achieve an average accuracy of 90.11% per image,which is an increase of 4.81% compared to the traditional Res Net-50 network.The main tasks completed in this research are as follows:(1)Research on pre-classification method of fire or smoke characteristics.Use the color features of the fire or smoke images to pre-classify the input images.If the pre-classification fails,it is judged that no fire has occurred,and there is no need to perform subsequent classifications,so as to eliminate the interference of some targets with similar colors of smoke and flames,such as various colors of other smoke,sunsets,and pedestrian clothing,as well as some targets with higher brightness but a certain distance from smoke and flames,such as clouds,spotlights and other targets.(2)Research on the dual attention mechanism method of channel attention and spatial attention.Aiming at the ambiguity of key information in images,channel and spatial attention mechanism are proposed.We use the channel attention mechanism module to extract the feature of the conceptual semantic information in the picture to solve ‘what’ in the image,smoke or fire and the spatial attention mechanism module to extract the feature of the spatial position information in the picture to solve ‘where’,the location of smoke or fire.(3)Research on bilinear pooling method.Use bilinear pooling to convert the outer product of two different vector spaces of channel and space into a new vector.This vector can be expressed as the semantic and interactive information of the two vectors in the same space.It is represented as the final vector,finally,the corresponding classification results of the image are output.(4)Smoke and fire image classification algorithm based on BAN experiment.Three different neural network models were selected as comparative experiments in the paper,including Resnet-50,Resnet-50(ADD)and Resnet-50(CAT).Take accuracy,precision,recall,and misclassification rate curves as evaluation indicators to test and analyze smoke images,flame images,and images without smoke and flame.The experimental results show that the smoke and fire image classification algorithm based on BAN proposed in this paper has achieved better results than the comparison model.