A multi-scale de-noising algorithm based on the convolution type of wavelet packet transformation was presented. This algorithm overcame shortcomings of the classical wavelet packet transformation, in which the length of sequences obtained always decreased by decomposition scales. The new algorithm improved estimated method of white noise standard deviation at each scale and thus kept the main edges of signal well. A new threshold function has been employed in this algorithm, which was simple in expression and as continuous as the Donoho's soft threshold function, and overcame the shortcoming of an invariable dispersion between the estimated wavelet coefficients and the decomposed wavelet coefficients of the soft-threshold method. Simulation results indicated that the new de-noising method suppressed the Pseudo-Gibbs phenomena near the singularities of the signal effectively and achieved better SNR gains than de-noising method based on classical wavelet packet transformation. 