Hidden Markov Tree Model of the Uniform Discrete Curvelet Transform Image for Denoising Yothin Rakvongthai
Introduction Curvelet Transform (Candes&Donoho 1999) Implementation: Fast Discrete Curvelet Transform (FDCT) (Candes et. al 2005) : in frequency domain Contourlet (Do&Vetterli 2005) : in time domain with wavelet-like tree structure Uniform Discrete Curvelet Transform (UDCT) (Nguyen&Chauris 2008) : in frequency domain with wavelet-like tree structure
Implementation
UDCT Marginal Statistics Kurtosis = 23.71 Kurtosis = 24.42 Kurtosis = E[(x-μ)4]/σ4 . Kurtosis of Gaussian = 3
Conditional Distribution (1) On parent (same position in next level) P(X|PX) Bow-tie shape uncorrelated but dependent
Conditional Distribution (2) On parent P(X|PX=px) Kurtosis=3.51 ~Gaussian
Hidden Markov Tree (HMT) Model Conditional distribution is Gaussian X depends on PX Use HMT to model the coefficients HMT model links between the hidden state variables of parent and children HMT parameters (parameters of the density function) can be trained using the expectation-minimization (EM) algorithm
Tree Structure of UDCT
HMT (1) c(j,k,n) – coefficient in scale j, direction k, position n S(j,k,n) – hidden state taking on values m = “S” or “L” with density function P(S(j,k,n)) Conditioned on S(j,k,n)=m, c(j,k,n) is Gaussian with mean μm(j,k,n) and variance σ2m(j,k,n) (m=Ssmall variance, m=Llarge variance)
HMT (2) The total pdf P(S(j,k,n)), μm(j,k,n), σ2m(j,k,n) can be trained from the EM algorithm (Crouse et al 1998). Define Θ = set of P(S(j,k,n)), μm(j,k,n), σ2m(j,k,n)
Denoising (1) Problem formulation: y = x+w ynoisy coefficients xdenoised coefficients wnoise coefficients with known variance Want to estimate x from the knowledge of y and variance of w
Denoising (2) Obtain Θ from EM algorithm The variance of denoised coefficients is
Denoising (3) The estimate of x
Denoising Results (1) PSNR = Peak Signal to Noise Ratio
Denoising Results (2) SSIM = Structure Similarity Index (Wang et. al 2004)
Denoising Results (3) Original Noisy (14.14dB) Wavelet (25.73dB) (SSIM 0.112) (SSIM 0.561) Contourlet (25.85dB) DT-CWT (26.54dB) UDCT (27.32dB) (SSIM 0.590) (SSIM 0.579) (SSIM 0.676)
Denoising Results (4) Original Noisy (14.14dB) Wavelet (23.38dB) (SSIM 0.184) (SSIM 0.508) Contourlet (22.94dB) DT-CWT (24.15dB) UDCT (24.35dB) (SSIM 0.479) (SSIM 0.557) (SSIM 0.570)
Denoising Results (5) Original Noisy (14.14dB) Wavelet (25.25dB) (SSIM 0.110) (SSIM 0.539) Contourlet (25.51dB) DT-CWT (25.99dB) UDCT (26.51dB) (SSIM 0.555) (SSIM 0.553) (SSIM 0.627)