function_mc_generateobservations.m

对图像进行local approximation处理的技术,应用于图像去模糊中.

function Output = GenerateObservations(Input,size_initial,v_support)

% function generates blurred noisy observations for Blind Deconvolution 
% problem in Multiframe (Multichannel) Imaging 
%
% Arguments:
%   Input.y - true images (y)
%   Input.v_true{:} - cell array of true PSFs
%   Input.BSNR(:) - array of BSNR value for noise
%   Input.shifts(:,:) - 2D array of vertical and horizontal shifts of
%                       observations
%   Input.NumberOfObservations - number of observations = min(length(Input.v_true),length(Input.BSNR))
%
%   size_initial=5; % support size of INITIAL PSF
%   v_support=zeros(15,15); % SUPPORT of PSFs, odd sizes only
%                           used for errors, illustration of PSFs
%
%
% Returns:
%   Output.v_init - initial PSFs
%   Output.v_true - true PSFs with same support
%   Output.sigma  - std of noise for each observation
%   Output.z      - array of observations
%   Output.Z      - FT of observations
%
% Observation model:
%   z(j) = y*v(j)+sigma(j)n
%   z(j) : blurred noisy observations, j=1,...,L
%   y    : true image (assumed as unknown)
%   v(j) : point-spread functions
%   n    : gaussian white noise with standard deviation sigma(j) for every 
%        jth obsrvation z(j)
%   *    : convolution
%
%
% The files are prepared in Tampere University of Technology, Institute of
% Information Technology, 2006.
%
% Vladimir Katkovnik, Dmitriy Paliy*
% *e-mail: [e-mail]

if nargin<3,
    v_support=zeros(15,15); % SUPPORT of PSFs, odd sizes only
    % used for errors, illustration of PSFs
end;

if nargin<2,
    size_initial=5; % support size of INITIAL PSF
end;

%%% FIX SEED FOR RANDOM PROCESSES  (OPTIONAL)
init=2055615788; randn('seed',init);

[yN,xN,zN] = size(Input.y);

Output.v = zeros(yN,xN,Input.NumberOfObservations,size(Input.y,3));

for i=1:Input.NumberOfObservations,
    %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
    %%%  GENERATES INITIAL PSF  %%%%%%%%%%%%%%%%%%%%%%%%%%
    %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%

    v = fspecial('gaussian',size_initial,1);
    v = function_KernelsForFFT( v, v_support);
    % pad PSF with zeros to whole image domain, and centers it.
    [ghy1,ghx1]=size(v);
    big_v1_initial=zeros(yN,xN); big_v1_initial(1:ghy1,1:ghx1) = v;
    Output.v_true{i} = function_KernelsForFFT( Input.v_true{i}, v_support);
        
    for ccolor=1:size(Input.y,3),
        Output.v_init(:,:,i,ccolor) = circshift(big_v1_initial,-round([(ghy1-1)/2 (ghx1-1)/2]));

        Output.v(:,:,i,ccolor) = Output.v_init(:,:,i,ccolor); % function_KernelsForFFT( Input.v_true{i}, zeros(yN,xN));
        
        % performs blurring (convolution is obtained by product in frequency domain)
        y_blur = real(ifft2(fft2(fftshift(function_KernelsForFFT( Input.v_true{i}, zeros(yN,xN)))).*fft2(Input.y(:,:,ccolor))));
        y_blur = circshift(y_blur,[Input.shifts(i,1) Input.shifts(i,2)]);

        %%%  ADDING NOISE to BLURRED SIGNAL  %%%%%%%%%%%%
        % sigma of noise in blurred image given the desired BSNR
        if Input.BSNR(i)>0,
            Output.sigma(i) = sqrt(norm(y_blur(:)-mean(y_blur(:)),2)^2 /(xN*yN*10^(Input.BSNR(i)/10)));
        else
            Output.sigma(i) = 0;
        end;

        n1 = Output.sigma(i)*randn(size(y_blur)); % white Gaussian noise with variance org_sigma^2
        Output.z(:,:,i,ccolor) = y_blur + n1;
        Output.Z(:,:,i,ccolor) = fft2(Output.z(:,:,i,ccolor));
    end;    
end;