A New Fire Detection Method Based On CNN-SVM

The occurrence of fire has brought a great threat to the human environment,and the performance of traditional detectors used to detect fires will be affected by various parameters of their own,so a more efficient fire detection method is urgently needed.As a commonly used algorithm for image processing,convolutional neural networks(CNN)have the advantage of automatically extracting image features,and VGG16,as a commonly used CNN architecture method,can deeply reflect the relationship between the depth of CNN and its performance,so VGG16 is selected as the CNN part of the new algorithm.The softmax classifier used for classification in the VGG16 model essentially converts the output of the CNN fully connected layer into probability,and does not have the ability to learn the model independently,so the generalization of softmax is not very high.As the most mature image classification tool in machine learning,support vector machine(SVM)can train and learn models based on images,which can enhance the robustness and generalization of the algorithm.Therefore,in order to be able to integrate the advantages of the two algorithms,we consider using SVM instead of the softmax classifier to make the model achieve better performance,and finally form the new fire detection method based on CNN-SVM in this paper.The flame image dataset used in this experiment is derived from the open source dataset of static fire images on https://collections.durham.ac.uk/files/r2d217qp536#.X5F5G2 gz Zn K.Through a large number of experimental comparisons,the results show that the new fire recognition algorithm is better than using one of the structures alone in balancing the relationship between the recognition accuracy and the recognition speed,and the new algorithm has higher robustness and generalization.