Study On Image Classification Algorithm And Application Based On Support Vector Machine

The automatic classification of massive images has become one of the research hot spots in many fields,including traditional methods such as remote sensing image detection,medical application,and face recognition.When operating on images,the highdimensional attributes of the image can make the classification result unsatisfactory,but with the development of technology in the field of machine learning,research on image classification is in full swing.Support Vector Machine(SVM)is used for image recognition,signal processing and regression prediction due to its excellent performance.However,the following problems still exist when using SVM to solve image classification:(1)First,it is the problem of the image itself.Generally speaking,the image is highdimensional data.If it is directly sent to the SVM classification,there is be a problem of curse of dimensionality.At the same time,many features of the image are unimportant.Removing unimportant features can not only solve the problem of dimension disaster,but also retain the important features of the image.(2)Second,the performance of SVM classification depends on the selection of kernel functions and kernel parameters.A suitable kernel function and optimized kernel parameters can greatly improve the classification performance of SVM.Therefore,in order to improve the performance of SVM in image classification,the main work of this master's is as follows.(1)The image is preprocessed to ensure that the input of SVM image feature information is important and fine.Specifically,we used image graying and linear interpolation algorithms to unify the type and size of the image.Then,in order to remove redundant features of the image and retain important information of the image,we used Convolutional Neural Network(CNN)performs feature extraction on the image.(2)For the selection of the kernel function,we choose the most commonly used RBF kernel function;and consider the selection of the kernel parameter as an optimization problem and use the Quasi-Bird Swarm Algorithm(QBSA)to solve this optimization problem while using the Markov chain analyzes the convergence of QBSA.Furthermore,we theoretically ensure the application of QBSA in the optimization of SVM parameters.In addition,12 benchmark functions are used to compare experiments with other representative algorithms.The experimental results verify the superiority of QBSA in optimization problems.Theoretical analysis and experimental results show that QBSA has a strong optimization ability,which can effectively prevent the optimization problem from falling into a local optimum.The optimized kernel parameters effectively improve the classification performance of SVM.(3)The image features extracted by CNN are sent to the SVM model optimized by QBSA parameters,and then the image features are classified.We call this image classification method CNN-QBSA-SVM,which implements the entire process from image feature extraction to classification.Finally,the hyperspectral data of Pavia University and the 8-class object images of the UCI dataset were used to analyze the classification performance of CNN-QBSA-SVM,and compared with other similar methods.The receiver operating characteristic(ROC)curve was used to evaluate the experimental results,which confirmed the superiority of CNN-QBSA-SVM in image classification.