1/9/12 11:36 AM

C:\MATLAB7\toolbox\images\images\deconvblind.m

1 of 9

function [J,P] = deconvblind(varargin) %DECONVBLIND Image restoration using blind deconvolution algorithm. % [J,PSF] = DECONVBLIND(I,INITPSF) deconvolves image I using maximum % likelihood algorithm, returning both deblurred image J and a restored % point-spread function PSF. The resulting PSF is a positive array of % the same size as the INITPSF, normalized so its sum adds to 1. The % PSF restoration is affected strongly by the size of its initial % guess, INITPSF, and less by its values (an array of ones is a safer % guess). % % To improve the restoration, additional parameters can be passed in % (use [] as a place holder if an intermediate parameter is unknown): % [J,PSF] = DECONVBLIND(I,INITPSF,NUMIT) % [J,PSF] = DECONVBLIND(I,INITPSF,NUMIT,DAMPAR) % [J,PSF] = DECONVBLIND(I,INITPSF,NUMIT,DAMPAR,WEIGHT) % [J,PSF] = DECONVBLIND(I,INITPSF,NUMIT,DAMPAR,WEIGHT,READOUT). % % Additional constraints on PSF can be provided via a user supplied % function: % [J,PSF] = DECONVBLIND(...,FUN,P1,P2,...,PN) % % FUN (optional) is a function describing additional constraints on the % PSF. There are four ways to specify FUN: as a function-handle, @, as % an inline object, or as a string containing either a function name or % a MATLAB expression. FUN is called at the end of each iteration. FUN % must accept the PSF as its first argument and may accept additional % parameters, P1, P2, ..., PN. FUN should return one argument, PSF, % that is the same size as the INITPSF, and satisfies the positivity % and normalization constraints. % % NUMIT (optional) is the number of iterations (default is 10). % % DAMPAR (optional) is an array that specifies the threshold deviation % of the resulting image from the image I (in terms of the standard % deviation of Poisson noise) below which the damping occurs. The % iterations are suppressed for the pixels that deviate within the % DAMPAR value from their original value. This suppresses the noise % generation in such pixels, preserving necessary image details % elsewhere. Default is 0 (no damping). % % WEIGHT (optional) is assigned to each pixel to reflect its recording % quality in the camera. A bad pixel is excluded from the solution by % assigning it zero weight value. Instead of giving a weight of one for % good pixels, you can adjust their weight according to the amount of % flat-field correction. Default is a unit array of the same size as % input image I. % % READOUT (optional) is an array (or a value) corresponding to the % additive noise (e.g., background, foreground noise) and the variance % of the read-out camera noise. READOUT has to be in the units of the % image. Default is 0. %

1/9/12 11:36 AM % % % % % % % % % % % % % % % % % % % % % % % % % % % % % % % % % % % % % % % % % % % % % % % % % %

C:\MATLAB7\toolbox\images\images\deconvblind.m

2 of 9

Note that the output image J could exhibit ringing introduced by the discrete Fourier transform used in the algorithm. To reduce the ringing use I = EDGETAPER(I,PSF) prior to calling DECONVBLIND. Note also that DECONVBLIND allows you to resume deconvolution starting from the results of an earlier DECONVBLIND run. To initiate this syntax, the input I and INITPSF have to be passed in as cell arrays, {I} and {INITPSF}. Then the output J and PSF become cell arrays and can be passed as the input arrays into the next DECONVBLIND call. The input cell array can contain one numeric array (on initial call), or four numeric arrays (when it is the output from a previous run of DECONVBLIND). The output J contains four elements, where J{1}=I, J{2} is the image resulted from the last iteration, J{3} is the image from one before last iteration, J{4} is an array used internally by the iterative algorithm. Class Support ------------I and INITPSF can be of class uint8, uint16, or double. DAMPAR READOUT have to be of the same class as the input image. Other have to be of class double. Output image J (or the first array output cell) is of the same class as the input image I. Output of class double. Example ------I = checkerboard(8); PSF = fspecial('gaussian',7,10); V = .0001; BlurredNoisy = imnoise(imfilter(I,PSF),'gaussian',0,V); WT = zeros(size(I));WT(5:end-4,5:end-4) = 1; INITPSF = ones(size(PSF)); FUN = inline('PSF + P1','PSF','P1'); [J P] = deconvblind(BlurredNoisy,INITPSF,20,10*sqrt(V),WT,FUN,0); subplot(221);imshow(BlurredNoisy); title('A = Blurred and Noisy'); subplot(222);imshow(PSF,[]); title('True PSF'); subplot(223);imshow(J); title('Deblured Image'); subplot(224);imshow(P,[]); title('Recovered PSF'); See also DECONVWNR, DECONVREG, DECONVLUCY, EDGETAPER, PADARRAY, PSF2OTF, OTF2PSF. Copyright 1993-2003 The MathWorks, Inc. $Revision: 1.4.4.3 $

and inputs of the PSF is

1/9/12 11:36 AM % % % % % % % % % % %

C:\MATLAB7\toolbox\images\images\deconvblind.m

3 of 9

References ---------"Acceleration of iterative image restoration algorithms, by D.S.C. Biggs and M. Andrews, Applied Optics, Vol. 36, No. 8, 1997. "Deconvolutions of Hubble Space Telescope Images and Spectra", R.J. Hanisch, R.L. White, and R.L. Gilliland. in "Deconvolution of Images and Spectra", Ed. P.A. Jansson, 2nd ed., Academic Press, CA, 1997. "Light Microscopic Images Reconstructed by Maximum Likelihood Deconvolution", Timothy J. Holmes et al. in "Handbook of Biological Confocal Microscopy", Ed. James B. Pawley, Plenum Press, New York, 1995

% Parse inputs to verify valid function calling syntaxes and arguments [J,P,NUMIT,DAMPAR,READOUT,WEIGHT,sizeI,classI,sizePSF,FunFcn,FunArg] = parse_inputs(varargin{:}); % 1. Prepare parameters for iterations % % Create indexes for image according to the sampling rate idx = repmat({ ':'},[1 length(sizeI)]); wI = max(WEIGHT.*(READOUT + J{1}),0); % at this point fw = fftn(WEIGHT); clear WEIGHT ; DAMPAR22 = (DAMPAR.^2)/2; % 2. L_R Iterations % lambda = 2*any(J{4}(:)~=0); for k = (lambda + 1) : (lambda + NUMIT), % 2.a Make an image and PSF predictions for the next iteration if k > 2, % image lambda = (J{4}(:,1).'*J{4}(:,2))/(J{4}(:,2).'*J{4}(:,2) + eps); lambda = max(min(lambda,1),0); % stability enforcement end Y = max(J{2} + lambda*(J{2} - J{3}),0); % image positivity constraint

...

- positivity constraint

if k > 2, % PSF lambda = (P{4}(:,1).'*P{4}(:,2))/(P{4}(:,2).'*P{4}(:,2) + eps); lambda = max(min(lambda,1),0); % stability enforcement end B = max(P{2} + lambda*(P{2} - P{3}),0); % PSF positivity constraint sumPSF = sum(B(:)); B = B/(sum(B(:)) + (sumPSF==0)*eps); % normalization is a necessary constraint, % because given only input image, the algorithm cannot even know how much % power is in the image vs PSF. Therefore, we force PSF to satisfy this % type of normalization: sum to one. % 2.b Make core for the LR estimation CC = corelucy(Y,psf2otf(B,sizeI),DAMPAR22,wI,READOUT,1,idx,[],[]);

1/9/12 11:36 AM

C:\MATLAB7\toolbox\images\images\deconvblind.m

4 of 9

% 2.c Determine next iteration image & apply positivity constraint J{3} = J{2}; H = psf2otf(P{2},sizeI); scale = real(ifftn(conj(H).*fw)) + sqrt(eps); J{2} = max(Y.*real(ifftn(conj(H).*CC))./scale,0); clear scale; J{4} = [J{2}(:)-Y(:) J{4}(:,1)]; clear Y H ; % 2.d Determine next iteration PSF & apply positivity constraint + normalization P{3} = P{2}; H = fftn(J{3}); scale = otf2psf(conj(H).*fw,sizePSF) + sqrt(eps); P{2} = max(B.*otf2psf(conj(H).*CC,sizePSF)./scale,0); clear CC H ; sumPSF = sum(P{2}(:)); P{2} = P{2}/(sumPSF + (sumPSF==0)*eps); if ~isempty(FunFcn), FunArg{1} = P{2}; P{2} = feval(FunFcn,FunArg{:}); end ; P{4} = [P{2}(:)-B(:) P{4}(:,1)]; end clear fw wI ; % 3. Convert the right array (for cell it is first array, for notcell it is % second array) to the original image class & output the whole thing num = 1 + strcmp(classI{1}, 'notcell' ); if ~strcmp(classI{2}, 'double' ), J{num} = changeclass(classI{2},J{num}); end if num == 2, % the input & output is NOT a cell P = P{2}; J = J{2}; end; %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%% % % Function: parse_inputs function [J,P,NUMIT,DAMPAR,READOUT,WEIGHT,sizeI,classI,sizePSF,FunFcn,FunArg] = parse_inputs(varargin) % % Outputs: % I=J{1} the input array (could be any numeric class, 2D, 3D) % P=P{1} the operator that distorts the ideal image % % Defaults: % NUMIT = [];NUMIT_d = 10; % Number of iterations, usually produces good

...

1/9/12 11:36 AM

C:\MATLAB7\toolbox\images\images\deconvblind.m

5 of 9

% result by 10. DAMPAR =[];DAMPAR_d = 0; % No damping is default WEIGHT =[];WEIGHT_d = 1; % All pixels are of equal quality, flat-field is one READOUT=[];READOUT_d= 0; % Zero readout noise or any other % back/fore/ground noise associated with CCD camera. % Or the Image is corrected already for this noise by user. FunFcn = '' ;FunFcn_d = ''; FunArg = {};FunArg_d = {}; funnum = [];funnum_d = nargin+1; checknargin(2,inf,nargin,mfilename); % no constraint on max number because of FUN args % First, assign the inputs starting with the cell/not cell image & PSF % switch iscell(varargin{1}) + iscell(varargin{2}), case 0, % no-cell array is used to do a single set of iterations classI{1} = 'notcell' ; J{1} = varargin{1}; % create a cell array in order to do the iterations P{1} = varargin{2}; case 1, msg = sprintf( 'In function %s, I and INITPSF must either both be cell arrays or both be numeric arrays.' ,... mfilename); eid = sprintf( 'Images:%s:IandInitpsfMustBeOfSameType' ,mfilename); error(eid,msg); case 2, % input cell is used to resume the interrupted iterations or classI{1} = 'cell' ; % to interrupt the iteration to resume them later J = varargin{1}; P = varargin{2}; if length(J)~=length(P), msg = sprintf( 'In function %s, I and INITPSF cell arrays must be of the same size.' , mfilename); eid = sprintf( 'Images:%s:IandInitpsfMustBeOfSameSize' ,mfilename); error(eid,msg); end end; % check the Image, which is the first array of the J-cell [sizeI, sizePSF] = padlength(size(J{1}), size(P{1})); classI{2} = class(J{1}); checkinput(J{1},{ 'uint8' 'uint16' 'double' },{ 'real' 'nonempty' 'finite' },mfilename, 'I',1); if prod(sizeI)<2, msg = sprintf( 'In function %s, input image must have at least two elements.' ,mfilename); eid = sprintf( 'Images:%s:inputImageMustHaveAtLeast2Elements' ,mfilename); error(eid,msg); elseif ~isa(J{1}, 'double' ), J{1} = im2double(J{1}); end % check the PSF, which is the first array of the P-cell

1/9/12 11:36 AM

C:\MATLAB7\toolbox\images\images\deconvblind.m

6 of 9

checkinput(P{1},{ 'uint8' 'uint16' 'double' },{ 'real' 'nonempty' 'finite' 'nonzero' }, mfilename,'INITPSF' ,1); if prod(sizePSF)<2, msg = sprintf( 'In function %s, initial PSF must have at least two elements.' ,mfilename); eid = sprintf( 'Images:%s:initPSFMustHaveAtLeast2Elements' ,mfilename); error(eid,msg); elseif ~isa(P{1}, 'double' ), P{1} = double(P{1}); end % now since the image&PSF are OK&double, we assign the rest of the J & P cells len = length(J); if len == 1, % J = {I} will be reassigned to J = {I,I,0,0} J{2} = J{1}; J{3} = 0; J{4}(prod(sizeI),2) = 0; P{2} = P{1}; P{3} = 0; P{4}(prod(sizePSF),2) = 0; elseif len ~= 4, % J = {I,J,Jm1,gk} has to have 4 or 1 arrays msg = sprintf( 'In function %s, the input cells must consist of 1 or 4 numerical arrays.' ,mfilename); eid = sprintf( 'Images:%s:inputCellsMustBe1or4ElementNumArrays' ,mfilename); error(eid,msg); else % check if J,Jm1,gk are double in the input cell if ~all([isa(J{2}, 'double' ),isa(J{3}, 'double' ),isa(J{4}, 'double' )]), msg = sprintf([ 'In function %s, second, third, and forth array of the input image cell' ... ' have to be of class double.' ],mfilename); eid = sprintf( 'Images:%s:ImageCellElementsMustBeDouble' ,mfilename); error(eid,msg); elseif ~all([isa(P{2}, 'double' ),isa(P{3}, 'double' ),isa(P{4}, 'double' )]), msg = sprintf([ 'In function %s, second, third, and forth array of the input PSF cell' ... ' have to be of class double.' ],mfilename); eid = sprintf( 'Images:%s:psfCellElementsMustBeDouble' ,mfilename); error(eid,msg); end end; % Second, Find out if we have a function to put additional constraints on the PSF % function_classes = { 'inline' , 'function_handle' , 'char' }; idx = []; for n = 3:nargin, idx = strmatch(class(varargin{n}),function_classes); if ~isempty(idx), [FunFcn,msg] = fcnchk(varargin{n}); %only works on char, making it inline if ~isempty(msg), if isstruct(msg)

1/9/12 11:36 AM

C:\MATLAB7\toolbox\images\images\deconvblind.m

7 of 9

eid = 'Images:deconvblind:invalidFun' ; error(eid, '%s', msg); else eid = sprintf( 'Images:%s:fcnchkError' , mfilename); error(eid,msg); end end FunArg = [{0},varargin(n+1:nargin)]; try % how this function works, just in case. feval(FunFcn,FunArg{:}); catch Ftype = { 'inline object' , 'function_handle' , 'expression ==>' }; Ffcnstr = { ' ' , ' ' ,varargin{n}}; msg = sprintf([ 'DECONVBLIND cannot continue because user supplied' ... ' %s %s\n failed with the error below.\n\n%s ' ], ... Ftype{idx},Ffcnstr{idx},lasterr); eid=sprintf( 'Images:%s:userSuppliedFcnFailed' ,mfilename); error(eid,msg) end funnum = n; break end end if isempty(idx), FunFcn = FunFcn_d; FunArg = FunArg_d; funnum = funnum_d; end % % Third, Assign the inputs for general deconvolution: % checknargin(3,7,funnum,mfilename); switch funnum, case 4, % deconvblind(I,PSF,NUMIT,fun,...) NUMIT = varargin{3}; case 5, % deconvblind(I,PSF,NUMIT,DAMPAR,fun,...) NUMIT = varargin{3}; DAMPAR = varargin{4}; case 6, % deconvblind(I,PSF,NUMIT,DAMPAR,WEIGHT,fun,...) NUMIT = varargin{3}; DAMPAR = varargin{4}; WEIGHT = varargin{5}; case 7, % deconvblind(I,PSF,NUMIT,DAMPAR,WEIGHT,READOUT,fun,...) NUMIT = varargin{3}; DAMPAR = varargin{4}; WEIGHT = varargin{5}; READOUT = varargin{6}; end % Forth, Check validity of the gen.conv. input parameters:

1/9/12 11:36 AM

C:\MATLAB7\toolbox\images\images\deconvblind.m

8 of 9

% % NUMIT check number of iterations if isempty(NUMIT), NUMIT = NUMIT_d; else %verify validity checkinput(NUMIT,{ 'double' },{ 'scalar' 'positive' 'integer' 'finite' }, mfilename,'NUMIT' ,3); end % DAMPAR check damping parameter if isempty(DAMPAR), DAMPAR = DAMPAR_d; elseif (numel(DAMPAR)~=1) && ~isequal(size(DAMPAR),sizeI), eid=sprintf( 'Images:%s:damparMustBeSizeOfInputImage' ,mfilename); error(eid, 'If not a scalar, DAMPAR has to be size of the input image.' ); elseif ~isa(DAMPAR,classI{2}), eid=sprintf( 'Images:%s:damparMustBeSameClassAsInputImage' ,mfilename); error(eid, 'DAMPAR has to be of the same class as the input image.' ); elseif ~strcmp(classI{2}, 'double' ), DAMPAR = im2double(DAMPAR); end if ~isfinite(DAMPAR), eid=sprintf( 'Images:%s:damparMustBeFinite' ,mfilename); error(eid, 'DAMPAR has to be finite.' ); end % WEIGHT check weighting if isempty(WEIGHT), WEIGHT = repmat(WEIGHT_d,sizeI); else numw = numel(WEIGHT); checkinput(WEIGHT,{ 'double' },{'finite' },mfilename, 'WEIGHT' ,5); if (numw ~= 1) && ~isequal(size(WEIGHT),sizeI), eid=sprintf( 'Images:%s:weightMustBeSizeOfInputImage' ,mfilename); error(eid, 'If not a scalar, WEIGHT has to have size of the input image.' ); elseif numw == 1, WEIGHT = repmat(WEIGHT,sizeI); end ; end % READOUT check read-out noise if isempty(READOUT), READOUT = READOUT_d; elseif (numel(READOUT)~=1) && ~isequal(size(READOUT),sizeI), eid=sprintf( 'Images:%s:readoutMustBeSizeOfInputImage' ,mfilename); error(eid, 'If not a scalar, READOUT has to be size of the input image.' ); elseif ~isa(READOUT,classI{2}), eid=sprintf( 'Images:%s:readoutMustBeSameClassAsInputImage' ,mfilename); error(eid, 'READOUT has to be of the same class as the input image.' ); elseif ~strcmp(classI{2}, 'double' ), READOUT = im2double(READOUT);

1/9/12 11:36 AM

C:\MATLAB7\toolbox\images\images\deconvblind.m

9 of 9

end if ~isfinite(READOUT), eid=sprintf( 'Images:%s:readoutMustBeFinite' ,mfilename); error(eid, 'READOUT has to be finite.' ); end;