Research And Application Of Virtual Sample Algorithm For Image Recognition In Machine Learning

With the development of science and technology,the image has a wide range of applications in social security,military security,information security,identity authentication,traffic supervision and so on.The image recognition technology is one of the key technologies of image application,and has received extensive attention.Especially in recent years,the application of machine learning algorithms in image recognition field has provided a new technical means for image classification,recognition and feature extraction,and has become a hotspot in this field.In order to solve the problem that machine learning is not satisfactory due to the lack of actual training samples,this thesis studies the influence about virtual samples' generation and number on machine learning performance.Firstly,this thesis introduced the basic principle of two representative machine learning algorithms,SVM and DBN.The mechanism and effect of feature extraction and dimension reduction are analyzed and verified experimentally.Secondly,this thesis presents a basic idea of evaluating the effectiveness of virtual samples based on feature level.The "mutual information","Euclidean distance" and the recognition accuracy in the feature evaluation are used as the evaluation of the validity of constructing virtual samples.At the same time,we propose and use "resampling","singular value reconstruction" and "contour reconstruction" to build a certain number of virtual samples to expand the original training set.Experiments verify that the sample data generated by the three virtual sample methods proposed in this thesis can effectively extend the data set and improve the recognition effect.Finally,the research results of the thesis are applied to SAR image and face image recognition,and experiments on the SVM and DBN verified the effect.Experiments on ORL,YALE and FERET data sets show that the recognition rate of face images can be improved effectively by extending virtual samples when the number of samples is not enough.In addition,the virtual sample is applied to the SAR image recognition for the first time.The experiment on the MSTAR data set shows that using virtual sample can obviously improve the classification result and have better recognition performance.The conclusion of this thesis proves the validity of the virtual sample in the machine learning algorithm and the wide application space.It also provides a powerful help for solving the small sample problem in the machine learning.