Abstract：present paper describes a new technique for stellar spectral recognition. Considering the characteristics of stellar spectral data, support vector machine (SVM) was adopted to build a recognition system as kernel. Because stellar spectral data sets are usually extremely noisy, the correct classification rate of direct applying SVM is low. Consequently, wavelet de-noising method was proposed to reduce noise first and extract the main characteristics of stellar spectra. Then SVM was used for the recognition. Based on the real-world stellar spectra contributed by Jacoby et al. (1984), it has proven that there will be a better performance using this composite classifier which combines wavelet and SVM than using SVM with principle component analysis data dimension reduction technique. From the experiment of comparison of discriminant analysis and SVM based on stellar spectra for evolutionary synthesis, we can see that the correct classification rate of SVM is higher than that of discriminant analysis methods, and a well generalization ability is achieved.
[1] GUO Ping, QIN Dong-mei, HU Zhan-yi(郭 平, 覃冬梅, 胡占义). Spectroscopy and Spectral Analysis(光谱学与光谱分析), 2003, 23(4): 811. [2] WANG Xi, XING Fei, GUO Ping. Proceedings of SPIE, 2003, 5286: 758. [3] Vapnik V. The Nature of Statistical Learning Theory. New York: Springer-Verlag, 1995. [4] Cortes C, Vapnik V. Machine Learning, 1995, 20: 273. [5] QIN Dong-mei, HU Zhan-yi, ZHAO Yong-heng(覃冬梅, 胡占义, 赵永恒). Spectroscopy and Spectral Analysis(光谱学与光谱分析), 2004, 24(4): 507. [6] Donoho D L. IEEE Transaction on Information Theory, 1995, 41: 613. [7] Jacoby G H, Hunter D A, Christian C A. J. Suppl. Ser.，1984, 56: 278. [8] Jolliffe I T, Principal Component Analysis, New York: Springer-Verlag, 1986.
