복합 노이즈 환경에서 개선된 노이즈 제거 알고리즘

Abstract

Noise, which is the main cause of degradation, takes different forms depending on its sources and types. Moreover, the AWGN and Impulse noise are representative. Therefore, a number of researches to remove noise and restore the original signal have been processed. However, the Fourier transform can not represent the local characteristics of signal since it transforms the whole signal by using a basis function. The STFT needs to have a proper window size for signal to analyze so has limited performance in the analysis for multiscale signals. The wavelet transform, which has been proposed to overcome limitations with the Fourier transform and STFT, has been applied to a variety of signal processing fields. In order to remove AWGN by using wavelets, Donoho and Johnstone proposed threshold-based algorithm. In the SSNF algorithm, noise was removed by using spatial correlation among wavelet coefficients in the neighboring scale. Additionally, UDWT was proposed to improve the noise removal performance of OWT. However, these existing noise removal algorithms do not have method removing noise mixed complexly. Therefore, in this thesis, in order to restore signal in complex noise environment, an algorithm using the inflection point of error distribution function, an algorithm using distribution characteristics based on the wavelet approximation coefficients and a noise removal algorithm by wavelet pair with half-sample delay characteristics were proposed. In a noise removal algorithm using the inflection point of error distribution function, first inflection point was established after peak point from smoothed data for histogram of error function and this preserves the edge part of signal and removes complex noise part concurrently. And a noise removal algorithm using distribution characteristics based on the wavelet approximation coefficients uses the average value of error function and normalized data number. According to characteristics of noisy signal the algorithm changed threshold value adaptively and separated the edge and noise parts. In a noise removal algorithm by wavelet pair with half-sample delay characteristics, the error data was acquired from difference between two approximation coefficients representing half-sample delay characteristics. And noise part was removed by applying new threshold value into the error data. In order to demonstrate the superiority of proposed algorithms in this paper, existing algorithms were compared by using SNR as the judgement criterion for noise removal.