Now a large telescope called LAMOST (Large Sky Area Multi-Object Fiber Spectroscopic Telescope) is under construction in the National Observatory in Beijing, China. After its scheduled completion in 2004, it is expected to collect more than 4,000 spectra in a single observation night. Such voluminous data demand automatic spectrum processing, and the previous mainly manual or interactive tèchniques become unsuitable. The main obstacles for automatic spectrum classification and recognition in LAMOST are two-fold: firstly, the spectra of celestial bodies are extremely noisy; and secondly, the spectra to be recognized are voluminous. Hence any acceptable technique must be both insensitive to noise and computationally efficient. This thesis is concentrated on the following two key issues in spectrum classification and recognition: redshift identification and automatic classification of spectra. The main results can be summarized as follows: 1. Redshift Identification by the pseudo-triangle method. Redshift is the most important parameter of celestial bodies outside the galaxy. Due to the coupling effect of redshift determination and spectrum classification, redshift identification is a key and very difficult problem in practice. A pseudo-triangle technique is proposed to deal with the problem, which, consists of the following three major steps: in the first step, the 3 wavelengths corresponding to the 3 highest intensity values of an unknown spectrum are selected to construct a pseudo-triangle, and the largest angle, being of independent of the unknown redshift value, of this triangle is calculated, and used as the index. In the second step, 3 model wavelengths are retrieved via this index from a pre-calculated look-up-table, which contains all the combinations of all the feature wavelengths of the model spectra. And finally based on the 3 corresponding wavelengths, the corresponding redshift value is derived. The main characteristics of our technique are its simplicity, efficiency and good robustness, and are demonstrated by experiments on simulated data as well as on real celestial spectra. 2. Atlas matching based on an RANSAC algorithm. The problem of atlas matching is in essence a point-matching problem from two point sets. It is equivalent to the point correspondence problem across two images in computer vision literature. The RANSAC based algorithms are widely used in computer vision due to its extremely good robustness. We have successfully applied an RANSAC based technique to the atlas-matching problem. Among the 960 available images, all but two incomplete images are correctly matched. Furthermore, the algorithm is quite computationally efficient, It takes no more than a few seconds for each matching. 3. Spectrum description language and classifying rule mining in spectra datasets using rough set We introduced a kind of new 'language' to describe a spectrum, which uses feature wavelengths, the corr
修改评论