Wavelet Image Denoising Based on Principle Component Analysis
Ting Rui · Journal of Chinese Computer Systems · 2006
Most of the existing methods on wavelet image denoising rely on accurate estimation of noise variance. In practice, however, the estimation of noise variance is very hard. To overcome this difficulty, this paper proposes a new method which utilizes noise energy, instead of its variance, to perform image denoising based on Principle Component Analysis (PCA) in the wavelet domain. First, wavelet decomposition is conducted on the noisy image, and PCA is used to extract local features; Second, the wavelet coefficients are reconstruct based on the top few principle components and the local noise energy is estimated based on the mean energy of reconstructed wavelet coefficients; Third, noise energy is subtracted from the original wavelet coefficients, which results in denoised wavelet coefficients; Finally, the inverse wavelet transform is performed to obtain the denoised image. A unique feature of the new algorithm is that it does not rely on the difficult task of noise variance estimation. It is therefore of great value in solving real-world problems. Compared with the commonly-used wavelet hard-threshholding and soft-thresholing methods, the new algorithm increases the PSNR by 2-8dB. Extensive experiments are conducted and the results demonstrate the superior denoising performance of the proposed algorithm.