隐马尔科夫树模型，用于计算图像的Contourlet系数

% pdfbdenoiseimagethmt.m % written by: Duncan Po % Date: January 14, 2004 % Clean the noisy image using the model and state probabilities of the % noisy image. The image is assumed to be square. % Usage: cleanimage = pdfbdenoiseimage( model, stateprob, nvar, nc, imname, imformat) % Inputs: model - model of the noisy image % stateprob - state probabilities of the noisy image % nvar - noise variance, normalized to the image range % (ie. should be between 0 and 1) % nc - noise variance in contourlet domain. Input '' if % unknown % imname - name of the image file % imformat - format of the image file (e.g. 'gif') % Output: cleanimage - the denoised image function cleanimage = pdfbdenoiseimage( model, stateprob, nvar, nc, imname, imformat) pyrfilter = '9-7'; dirfilter = 'pkva'; nlevel = model{1}.nlevels; for lev = 1:nlevel levndir(lev) = log2(length(model{1}.stdv{lev})*length(model)); end; if nargin == 6 coef = contourlet(pyrfilter, dirfilter, levndir, imname, imformat); elseif nargin == 5 coef = contourlet(pyrfilter, dirfilter, levndir, imname); end; % Verify image is square and obtain image dimensions [nrow, ncol] = size(coef{2}{1}); imagestats = imfinfo(imname, imformat); imdim = imagestats.Width; if imdim ~= imagestats.Height error('Image must be square.'); end for state = 1:size(stateprob{1}{1},2) for lev = 1:nlevel+1 newstateprob{state}{lev} = []; covariance{state}{lev} = []; end; end; % process stateprob, and covariance so that it has the same structure as coef for state = 1:size(stateprob{1}{1},2) for dir = 1:2.^levndir(1) for scale = 1:length(stateprob{dir}) tempsp{scale} = stateprob{dir}{scale}(:,state); end; tempstateprob = tree2contourlet(tempsp, dir, levndir, nrow, ncol); for lev = 1:nlevel newstateprob{state}{lev+1} = [newstateprob{state}{lev+1} tempstateprob{lev}]; for subdir = 1:length(model{dir}.stdv{lev}) covariance{state}{lev+1} = [covariance{state}{lev+1} ... {ones(size(coef{lev+1}{(dir-1)*(2.^(levndir(lev)-levndir(1)))+subdir},1),... size(coef{lev+1}{(dir-1)*(2.^(levndir(lev)-levndir(1)))+subdir},2))*... model{dir}.stdv{lev}{subdir}(state)*model{dir}.stdv{lev}{subdir}(state)}]; end; end; end; end; % calculates the noise variance in contourlet domain if isempty(nc) nc = contournc(nvar, pyrfilter, dirfilter, levndir, imdim); end; % computes the attenuator for each coefficient and state for state = 1:length(covariance) for lev = 2:nlevel+1 for dir = 1:length(covariance{state}{lev}) [covrow, covcol] = size(covariance{state}{lev}{dir}); covariance{state}{lev}{dir}=max(covariance{state}{lev}{dir}... -nc{lev}{dir}, zeros(covrow, covcol))./covariance{state}{lev}{dir}; end; end; end; % multiplies each contourlet coefficient with its corresponding state probabilities for state = 1:length(newstateprob) for lev = 2:nlevel+1 for dir = 1:length(newstateprob{state}{lev}) newcoef{state}{lev}{dir} = newstateprob{state}{lev}{dir}... .*coef{lev}{dir}; end; end; end; % we don't denoise the scaling function since its SNR is high cleancoef{1} = coef{1}; % initialize the clean coefficients for lev = 2:nlevel+1 for dir = 1:length(newstateprob{state}{lev}) cleancoef{lev}{dir} = zeros(size(coef{lev}{dir},1), size(coef{lev}{dir},2)); end; end; % denoise for state = 1:length(newstateprob) for lev = 2:nlevel+1 for dir = 1:length(newstateprob{state}{lev}) cleancoef{lev}{dir}=cleancoef{lev}{dir}+covariance{state}{lev}{dir}... .*newcoef{state}{lev}{dir}; end; end; end; cleanimage = pdfbrec(cleancoef, pyrfilter, dirfilter); figure; imshow(uint8(cleanimage));