Projection-free kernel principal component analysis for denoising

NEUROCOMPUTING

卷:357

Article

Anh Tuan Bui¹  Im, Joon-Ku²  Apley, Daniel W.¹  Runger, George C.³ 

[1] Northwestern Univ, Dept Ind Engn & Management Sci, 2145 Sheridan Rd, Evanston, IL 60208 USA

[2] Anthem Inc, 233 South Wacker Dr,Suite 3700, Chicago, IL 60606 USA

[3] Arizona State Univ, Sch Comp Informat & Decis Syst Engn, 699 S Mill Ave, Tempe, AZ 85281 USA

关键词: Image processing;    Feature space;    Pattern recognition;    Preimage problem;   

DOI  :  10.1016/j.neucom.2019.04.042

来源: Elsevier

【 摘 要 】

Kernel principal component analysis (KPCA) forms the basis for a class of methods commonly used for denoising a set of multivariate observations. Most KPCA algorithms involve two steps: projection and preimage approximation. We argue that this two-step procedure can be inefficient and result in poor denoising. We propose an alternative projection-free KPCA denoising approach that does not involve the usual projection and subsequent preimage approximation steps. In order to denoise an observation, our approach performs a single line search along the gradient descent direction of the squared projection error. The rationale is that this moves an observation towards the underlying manifold that represents the noiseless data in the most direct manner possible. We demonstrate that the approach is simple, computationally efficient, robust, and sometimes provides substantially better denoising than the standard KPCA algorithm. (C) 2019 Elsevier B.V. All rights reserved.

【 授权许可】

Free