ℓ1-Regression based subdivision schemes for noisy data. (January 2015)

• Record Type: Journal Article
• Authors: Mustafa, Ghulam; Li, Hao; Zhang, Juyong; Deng, Jiansong
• Is Part Of: Computer aided design. Volume 58(2015)
• Page Start: 189
• Page End: 199
• Publication Date: 2015-01
• Subjects: ℓ1-Regression -- Data restoration -- Outlier detection -- Subdivision scheme -- Iterative reweighted least squares; Computer-aided design -- Periodicals; Engineering design -- Data processing -- Periodicals; Computer graphics -- Periodicals; Conception technique -- Informatique -- Périodiques; Infographie -- Périodiques; Computer graphics; Engineering design -- Data processing; Periodicals; Electronic journals; 620.00420285
• DOI: 10.1016/j.cad.2014.08.003
• Languages: English
• ISSNs: 0010-4485

Abstract

Abstract: Fitting curve and surface by least-regression is quite common in many scientific fields. It, however cannot properly handle noisy data with impulsive noises and outliers. In this article, we study ℓ 1 -regression and its associated reweighted least squares for data restoration. Unlike most existing work, we propose the ℓ 1 -regression based subdivision schemes to handle this problem. In addition, we propose fast numerical optimization method: dynamic iterative reweighted least squares to solve this problem, which has closed form solution for each iteration. The most advantage of the proposed method is that it removes noises and outliers without any prior information about the input data. It also extends the least square regression based subdivision schemes from the fitting of a curve to the set of observations in 2-dimensional space to a p -dimensional hyperplane to a set of point observations in ( p + 1 ) -dimensional space. Wide-ranging experiments have been carried out to check the usability and practicality of this new framework. Highlights: An ℓ 1 -regression based subdivision scheme is proposed to handle noisy curve/surface data with outliers. A fast numerical optimization method named dynamic iterative reweighted least squares is proposed to solve this problem. The most advantage of the proposed method is that it removes noises and outliers without any prior information about the input data.