Research On SVDD And OCSVM For Pattern Denoising

Kernel methods(KMs)are one of the most important algorithms in Pattern Recognition and are extensively learned and applied.Support vector data description(SVDD)and one-class support vector machine(OCSVM)are the KMs of studying the problem of one-class classification and applying for pattern denoising.However,SVDD-based pattern denoising needs many iterations to obtain the final denoised sample.Not only does each iteration spend much time on loops and matrix calculations,but also it have an negative impact on algorithm efficiency when the data is huge.Then,there have a profound effect on the effectiveness of pattern denoising because it doesn't take the data distribution into consideration when building up the hyper-plane in OCSVM.Therefore,we explore the two problems in the article and the content described as following:With respect to the problem of SVDD-based pattern denoising,the result of each iteration is that the denoised sample moves a little close to the classification boundary.So,combining the first denoised sample with the noise sample,we propose a new method to improve the iterations on pattern denoising by using the mathematical method and strategy to work out the formula of the denoised sample.Experiments on toy datasets and real-world datasets demonstrate the feasibility of the algorithm and we make both comparisons and analysis on time efficiency.Comparing with other denoising methods,the proposed method have a better performance in the case of similar denoising effect;Since OCSVM doesn't consider the data distribution and establish the unreasonable hyper-plane.,in the case of Gaussian kernel,we bring the data distribution into OCSVM which is equivalent to SVDD and propose a new method based on data distribution with OCSVM for pattern denoising.In contrast with SVDD-based pattern denoising,the proposed algorithm takes the existing form of the samples in data space into full consideration and the classification hyper-plane we find in feature space is reasonable.The data distribution is important for improving the effect of algorithm and establishing the hyper-plane.The results of experiments on handwritten digit datasets prove out this point.According to the explorations in the thesis,the results of experiments demonstrate the efficiency and feasibility of the algorithm and obtain an inspiring performance.It lays the foundation for the better study of pattern denoising.