The authors propose a novel method of feature extraction for stellar spectra parameterization. The basic procedures are: First, stellar spectra are decomposed by multi-scale Harr wavelet and the coefficients with high-frequency are rejected. Secondly, the optimal features are detected by the lasso algorithm. Finally, we input the optimal feature vector to non-parametric regression model to estimate the atmospheric parameters. Haar wavelet can remove the high-frequency noise from the stellar spectrum. Lasso algorithm can further compress data by analyzing their significance on parameterization and removing redundancy. Experiments show that the proposed Haar+lasso method improves the accuracy and efficiency of the estimation. The authors used this scheme to estimate the atmospheric parameters from a subsample of some 40 000 stellar spectra from SDSS. The accuracies of our predictions (mean absolute errors) for each parameter are 0.007 1 dex for log Teff, 0.225 2 dex for log g, and 0.199 6 dex for ［Fe/H］. Compared with the results of the existing literature, this scheme can derive more accurate atmospheric parameters.