Abstract: Noise of signal not only reduces the quality of signal but also interferes the validity of correlative arithmetic seriously. Therefore, effective and robust estimation of noise variance is very important for various signal processing. A new method is proposed to estimate noise variance. A Gaussian mixture model (GMM) is used to model the high frequency wavelet coefficients (HFWC). The parameters of the mixture model are obtained with the EM iterative algorithm. The HFWC will be classified as noises class and edges class in the GMM when the parameters meet a certain condition. Based on the correlation among HFWC, GMM is used again to classify the coefficients of the noise as well as to take the noise estimation. Finally, the variance of noise signals is calculated and regarded as the noise variance estimation of original signal. Based on the estimation algorithm, wavelet threshold denoising is applied to reverse engineering. The denoising effect of practical signal is perfect in smoothness and feature preserving. The examination indicates that this estimation method of noise variance has certain adaptability to different noise, moreover, the denoising method of wavelet threshold can be simply achieved and applied in most situations.

Key words: Correlation of the coefficients, Estimation of noise variance, Gaussian mixture model, Wavelet threshold denoising