Research On Kernel Method For Support Vector Machine And Its Application In Forest Fire Video Recognition

The SVM (Support Vector Machine, SVM) is a kind of machine learning model which was put forward and developed in1995by AT&TBell laboratory research team led by Vapnik. The SVM is another representative mathematical model in the field of machine learning after the neural network model, and it has unique advantages in solving the problems of high dimension, small samples and nonlinear pattern recognition. Now it has been applied in many areas of machine learning and pattern recognition.As a new mathematical model, the SVM has the extreme development potential. Among these, the research of selection and the construction method of kernel function is the focus of support vector machine (SVM) research. The basic function of kernel function is to transform the two input vectors in lower dimension and calculate the inner product value in the high dimensional space, and the result is that two groups of linear and undivided samples which were in low dimensional space could become linearly separable when they are mapped to high-dimensional space. So the selection of kernel function plays a decisive role in the classification performance of support vector machine. But, so far, there is still not a guiding theory of how to select and construct the kernel function.This article puts forward the kernel function generalization ability evaluation model and dynamic width and scale Gaussian kernel function model, kernel matrix model based on Gaussian kernel multi-core model and the similarity by analyzing the support vector machine and kernel function theory. It carries out the simulation analysis for each model and uses the forest fire video identification to compare the advantages and disadvantages and the applicable occasions of each model in the end. The main work includes:1. The kernel function evaluation model and evaluation method studyBy discussing the commonly used evaluation methods, the article puts forward an improved evaluation model and the principle of optimizing AUC based on the AUC assessment criteria, and uses the method of experimental statistics to compare the commonly used model standards with improved model standards, providing the performance evaluation of commonly used kernel function.2. The dynamic kernel width and dynamic kernel scale research of the Gaussian kernel functionAccording to specific classification problem, the article dynamically adjusts the width of the Gaussian kernel function, so as to prevent the phenomenon of too much learning in area with larger data density and the phenomenon of less learning in area with smaller data density. At the same time, for the area with larger data density, the article uses the measures of small scale, and for the area with smaller data density, the article uses the measures of big scale, but the article keeps the same data relationship to get a more accurate description of data measurement, making the accuracy of classification is improved. Finally according to the above two designs, the article improves the Gaussian kernel function based on dynamic kernel width and dynamic kernel scale.3. The multikernel function research based on the improvement of Gaussian kernel functionThe article introduces a concept of the pyramid, and uses many traditional Gaussian kernel functions and dynamic width and scale kernel functions to construct the multikernel function. According to the sample density of the different layers of the pyramid, it respectively chooses different parameters and different types of kernel functions, and under the premise of retaining the advantages of Gaussian kernel function, it solves the defects of approximation and assumption which are brought by single kernel function, and reduces the risk of local generalization.4. The kernel matrix construction method research based on the similarityAccording to the definition of kernel function and the basic kernel matrix which is constructed by the original samples and data, the article uses the scattered data interpolation method to determine the inner product value of the interest points in feature space, so as to construct the extensive kernel matrix which satisfies the Mercer condition. It mainly involves the judge of kernel matrix, the discussion of the advantages and disadvantages of different construction methods of basic kernel matrix and extensive kernel matrix.5. New kernel method application in forest fire video recognitionMade suitable for forest fire video recognition feature selection method, for improving the Gaussian kernel function, multi-core functions and nuclear matrix actual validation study, comparing its classification performance, generalization and classification efficiency.