研究了天体光谱的特征提取问题, 这是光谱自动处理中的一个关键环节。 通过特征提取, 不仅能够约简数据、 减少冗余, 而且亦能抑制噪声干扰, 对识别系统的精度和效率均有重要影响。 提出了一种基于空间转换和分解的特征分析模型(STP), 基于此, 可实现对常用光谱特征提取方法的分析, 例如, 无监督的主成分分析(PCA), 小波变换(Wavelet), 有监督的支持向量机(SVM), 相关向量机(RVM)和线性判别分析方法(LDA)等。 在STP模型中, 关注的核心要素是特征提取中对数据成分的分解、 重组, 以及噪声的抑制和冗余的消除。 亦在STP框架的基础上, 给出了一种逻辑和实现均较为简单的特征提取方法: 基于曲线拟合与下采样的光谱特征提取(EFCD)。 研究的一个重要发现是, 在一些分类问题中文献中设计巧妙的特征提取方法并不一定是决定性的: 即使采用通常的信号下采样方法提取特征, 亦能获得良好的光谱识别性能, 而重要的仅仅是需要将特征数量保持在一定的水平以上即可。 研究中, 选用的测试数据是SDSS中的Galaxy和QSO两类河外天体实测光谱, 他们一般具有较大的红移, 在天体光谱识别中具有较强的代表性。

The present focuses on the celestial spectra feature extraction problem, which is a key procedure in automatic spectra classification. By extracting features, the authors can reduce redundancy, alleviate noise influence, and improve accuracy and efficiency in spectra classification. The authors introduced a novel feature analysis framework STP (space transformation and partition), which focuses on four essential components in feature extraction: decompose and reorganize spectrum components, reorganize, alleviate noise influence and eliminate redundancy. Based on STP, we can analyze most of the available feature extraction methods, for example, the unsupervised methods principal component analysis (PCA), wavelet transform, the supervised methods support vector machine (SVM), relevance vector machine (RVM), linear discriminant analysis (LDA), etc. We introduced a novel feature analysis framework and proposed a novel feature extraction method. The outstanding characteristics of the proposed method are its simplicity and efficiency. Researches show that it is sufficient to extract features by the proposed method in some cases, and it is not necessary to use the sophisticated methods, which is usually more complex in computation. The proposed method is evaluated in classifying Galaxy and QSO spectra, which is disturbed by red shift and is representative in automatic spectra classification research. The results are practical and helpful to gain novel insight into the traditional feature extraction methods and design more efficient spectrum classification method.