RN-SM：Stellar Spectral Classification Algorithm Based on ResNet
Feature Extraction
YANG Jia-ming1, TU Liang-ping1, 2*, LI Jian-xi2, MIAO Jia-wei1
1. School of Electronic and Information Engineering, University of Science and Technology, Anshan 114051，China
2. School of Mathematics and Statistics, Minnan Normal University, Zhangzhou 363000, China
Abstract：The advent of the Large Survey Telescope made it possible to build spectral databases of stellar. To study the massive stellar spectrum data in the database more efficiently and effectively, it is necessary to develop a fast and efficient automatic stellar spectrum processing algorithm. Based on the deep learning model ResNet, a hybrid deep learning algorithm named RN-SM is built in this paper. The algorithm consists of five steps: ① Normalization processing: A linear normalization function is used to normalize the stellar spectrum, ensuring it has a uniform scale. ② Denoising processing: The Ces algorithm is used to denoise the stellar spectrum, removing noise from the data. ③ Composite RGB image: Three channels of an RGB image, corresponding to the gray image generated by the same spectrum. The superposition of the same spectrum makes the main features of the stellar spectrum more pronounced and easier to work with in subsequent analysis. Here, we normalize the continuous spectrum of the stellar spectrum so that the content shown in the RGB image is the spectral line information of the stellar spectrum. At the same time, we analyze the feasibility of data conversion (synthetic RGB image) by using the main spectral line information of the stellar spectrum as a reference and investigating whether the relevant pixel position of the synthetic RGB image contains these features. It is proven that the method of data conversion (synthesizing an RGB image) proposed in this paper is feasible. ④ Feature extraction: To facilitate the connection of the SoftMax algorithm, the ResNet algorithm was used to extract features from stellar spectra. The 1×2 048 feature vector from the 64×64 RGB image was extracted. The ResNet algorithm contains 49 convolution layers in total. Automatic classification: The feature vector is transferred to the SoftMax module for automatic classification. The loss function used by SoftMax is the sum of the dataset loss and the regularization term loss. After 10000 iterations, the loss function becomes stable. When the RN-SM algorithm uses the spectra of A, B, dM, F, G, gM, and K-type stars with R-band signal-to-noise ratio greater than 30 for classification, the classification accuracy is 0.91. This classification accuracy is also higher than that of the CNN+Bayes, CNN+Knn, CNN+SVM, CNN+AdaBoost, and CNN+RF algorithms, at 0.862, 0.876, 0.894, 0.868, and 0.889, respectively.
杨佳铭，屠良平，李建喜，苗嘉伟. RN-SM：基于ResNet特征提取的恒星光谱分类算法[J]. 光谱学与光谱分析, 2025, 45(06): 1670-1679.
YANG Jia-ming, TU Liang-ping, LI Jian-xi, MIAO Jia-wei. RN-SM：Stellar Spectral Classification Algorithm Based on ResNet
Feature Extraction. SPECTROSCOPY AND SPECTRAL ANALYSIS, 2025, 45(06): 1670-1679.
[1] Liu C, Cui W, Zhang B, et al. Research in Astronomy and Astrophysics, 2015, 15(8): 1137.
[2] Li X, Lin Y, Qiu K. Research in Astronomy and Astrophysics, 2019, 19(8): 111.
[3] Wang Ke, Guo Ping，Luo A-li. Monthly Notices of the Royal Astronomical Society, 2017, 465(4): 4311.
[4] Mese G D, López-Cruz O, Schuster W J, et al. Monthly Notices of the Royal Astronomical Society, 2020, 494(2): 2995.
[5] Courtney L Crawford, Patrick Tisserand, Geoffrey C Clayton, et al. Monthly Notices of the Royal Astronomical Society, 2023, 521(2): 1674.
[6] Vincent O, Bergeron P, Dufour P. Monthly Notices of the Royal Astronomical Society, 2023, 521(1): 760.
[7] Bu Yude, Pan Jingchang, Jiang Bin, et al. Publications of the Astronomical Sciety of Japan, 2013, 65(4): 81.
[8] Wang Lili, Zheng Wenyan, Rong Lixia, et al. New Astronomy, 2023, 99: 101965.
[9] Du Changde, Luo Ali, Yang Haifeng. New Astronomy, 2017, 51: 51.
[10] Zhang Jingmin, Ma Chenye, Wang Lu, et al. Chinese Astronomy and Astrophysics, 2020, 44(3): 334.
[11] FAN Ya-wen, LIU Yan-ping, QIU Bo, et al（范雅雯，刘艳萍，邱 波, 等）. Spectroscopy and Spectral Analysis(光谱学与光谱分析), 2023, 43(8): 2523.
[12] Xue Jianqiao, Li Qibin，Zhao Yongheng. Chinese Astronomy and Astrophysics, 2001, 25(1): 120.
[13] Yang Jiaming, Tu Liangping, Liu Hao, et al. Current Science, 2024, 127(11): 1301.
