A novel scheme SVR(Haar) is proposed in the present work for automatically estimating the physical parameters of stellar spectra. The observed spectrum is disturbed usually by noise which is caused by the universe radiation, the atmosphere and observation equipment. Furthermore, the noise usually is the component of the spectrum with higher frequency. Therefore, we propose to extract features with Haar wavelet by removing higher frequency components. Researches show that this procedure can improve the accuracy of the estimation. Secondly, the support vector regression model is employed for estimating physical parameters of the stellar spectra. In this method, the ε insensitive domain techniques can further improve the probability to the slight distortion of the spectrum from imperfect calibration, and enhance the robustness of the proposed scheme. To check the effectiveness of the proposed scheme SVR(Haar), we did experiments extensively on authoritative simulated stellar spectra and real spectra observed by SLOAN, and compared it with the typical methods in the literature. The results show that the SVR(Haar) is better than the principal component analysis and non-parametric regression model in the literature.