In order to extract incipient weak features of rolling bearing from strong background noise, a new approach based on blind deconvolution and improved spectral subtraction method (SSM) is proposed. Firstly, a convolution model of rolling bearing vibration signal is presented. The transfer process of impulses is described, and a blind deconvolution filter is optimized to detect the impulsive components by means of the higher-order statistical (HOS) properties. Then, an effective methodology-SSM is introduced to differentiate noise from impulses to enhance the ability of features. To reduce additional noise induced by random impulse excitations, using modified parameters, an improved spectral subtraction is proposed through fault information existing in low and high frequency modulation areas. Furthermore, residual noise energy and signal distortion criterion are employed to evaluate the parameters. Test signal analysis shows the superiority of improved SSM. Finally, experimental data are used to verify the excellent reliability of blind deconvolution along with the improved SSM.