Research On Flame Recognition Algorithm Based On Machine Learning

With the development of civilization,the threat of fire to the safety of human life and property keeps a higher potential growth.Traditional fire detection mainly relies on sensor technology,which judges a fire on the changes of light intensity and temperature.However,sensors generally have the problems of limited range and generalization in different scenes.In contrast,the flame recognition algorithm based on image technology and machine learning has the advantages of high recognition accuracy,strong anti-interference and the ability of generalization.With the help of deep learning and neural network,it can extract two-dimensional features to achieve the ideal training results.In this dissertation,a soft decision voting classifier based on ensemble learning,a pseudo label multi-layer perceptron based on semi supervised learning and a ladder network algorithm are proposed.The results show that the former effectively overcomes the low efficiency of traditional machine learning algorithm,while the latter reduces the labor cost of data set annotation.The dissertation is organized as following:In Chapter 1,the background and the significance of research,the arts of flame recognition are stated,and the brief introduction of main content and structure of this dissertation are given in this chapter at the same time.Chapter 2 first introduces the basic theory of machine learning and a variety of machine learning algorithms,then it analyzes the feature extraction of flame recognition,introduces the Histogram of Oriented Gradient(HOG)to extract features,explains the training process and then verifies the learning performance of four machine learning algorithms based on the flame data set.In Chapter 3,the enhancement method of traditional machine learning algorithm is introduced.With various paths of ensemble learning brought in.,it explains the classifier enhancement principle of ensemble learning combined with specific algorithms.At last,an ensemble learning model based on soft voting decision is proposed.Combined with a variety of traditional machine learning classifiers,verifying the classification enhancement ability of ensemble learning is successful.In Chapter 4,firstly,based on the weakness of traditional machine learning,deep learning is introduced.The birth and development of neural network,and the deep feedforward network and back propagation algorithm is explained in seqution and details.On this basis,the layered architecture of convolutional neural network(CNN),the role and the functionof each layer is analyzed.Finally,the self built convolutional neural network and mature model Alexnet are used to verify the performance of the model,using the aspects of tdistribution random domain embedding and loss function.In Chapter 5,from the aspect of high tagging cost in supervised learning,semi supervised learning is introduced.Starting from the field of medical image and text research,the classification and basic principles of semi supervised learning are analyzed.Combined with the characteristics of deep learning which can improve the ability of feature learning,two semi-supervised deep learning theories,pseudo label multi-layer perceptron and ladder network,are proposed.And it is proved that semi supervised deep learning has a good performance in less sample labeling cost and higher accuracy.Chapter 6 is the summary and prospect of research.According to the inherent advantages and disadvantages of machine learning and deep learning,as well as the comprehensive technical points which are not considered in this paper,the future research work is planned and prospected.