How to output a struct array when I use the nlfilt

I have the following function which calculates statistic parameters. I would like to pass this function to nlfilter to do the calculation for a whole image. But the output of nlfilter must be a scalar.

How can I convert this to a function handle suitable for use with nlfilter so I can save the output of the function getStatistics2?

The getStatistics2 function's output is an struct array.

function [out] = getStatistics2(D)
D = double(D);   
% out.MAX = max(D);%maximum
% out.MIN = min(D);%minimum
out.MEA = mean(D);%mean
out.MAD = mad(D);% mean absolute deviation y=mean(abs(X-mean(x)))
out.MED = median(D);%median
out.RAN = max(D) - min(D);%range
out.RMS = rms(D);%root mean square
out.STD = std(D);%stardard deviation
out.VAR= var(D);%variance

This is an interesting question. What's interesting is that your approach is almost perfect. The only reason it fails is because struct cannot be constructed using a numeric scalar input (i.e. struct(3)). The reason I mention this is because somewhere during the execution of nlfilter (specifically in mkconstarray.m), it calls the the following code:

repmat(feval(class, value), size);

Where:

class is 'struct'.
value is 0.
size is the size() of the input image, e.g. [100,100].

... and this fails because feval('struct', 0), which is equivalent to struct(0) - and this we already know to be invalid.

So what do we do? Create a custom class that can be constructed this way!

Here's an example of one such class:

classdef MyStatsClass % Value class

  properties (GetAccess = public, SetAccess = private)
    MAX@double scalar = NaN;  % Maximum
    MIN@double scalar = NaN;  % Minimum
    MEA@double scalar = NaN;  % Mean
    MAD@double scalar = NaN;  % Mean absolute deviation y = mean(abs(X-mean(x)))
    MED@double scalar = NaN;  % Median
    RMS@double scalar = NaN;  % Root mean square
    STD@double scalar = NaN;  % Stardard deviation
    VAR@double scalar = NaN;  % Variance
    RAN@double scalar = NaN;  % Range    
  end % properties

  methods (Access = public)
    %% Constructor:
    function obj = MyStatsClass(vec)
      %% Special case:
      if (nargin == 0) || (numel(vec) == 1) && (vec == 0)
        % This happens during nlfilter allocation
        return
      end      
      %% Regular case:
      obj.MAX = max(vec(:));       
      obj.MIN = min(vec(:));
      obj.MEA = mean(vec(:));
      obj.MAD = mad(vec(:));
      obj.MED = median(vec(:));
      obj.RMS = rms(vec(:));
      obj.STD = std(vec(:));
      obj.VAR = var(vec(:));
      obj.RAN = obj.MAX - obj.MIN;
    end % default constructor
  end % public methods
end % classdef

And here's how you can use it:

function imF = q35693068(outputAsStruct)
if nargin == 0 || ~islogical(outputAsStruct) || ~isscalar(outputAsStruct)
  outputAsStruct = false;
end

rng(35693068); % Set the random seed, for repeatability
WINDOW_SZ = 3;
im = randn(100);
imF = nlfilter(im, [WINDOW_SZ WINDOW_SZ], @MyStatsClass);

% If output is strictly needed as a struct:
if outputAsStruct
  warning off MATLAB:structOnObject
  imF = arrayfun(@struct,imF);
  warning on MATLAB:structOnObject
end

Notice that I have added an optional input (outputAsStruct) that can force the output to be a struct array (and not an array of the type of our custom class, which is functionally identical to a read-only struct).

Notice also that by default nlfilter pads your array with zeros, which means that the (1,1) output will operate on an array that looks like this (assuming WINDOW_SZ=3):

[0    0      0    
 0  1.8096 0.5189 
 0 -0.3434 0.6586]

and not on im(1:WINDOW_SZ,1:WINDOW_SZ) which is:

[ 1.8096 0.5189 0.2811
 -0.3434 0.6586 0.8919
 -0.1525 0.7549 0.4497]

the "expected result" for im(1:WINDOW_SZ,1:WINDOW_SZ) will be found further "inside" the output array (in the case of WINDOW_SZ=3 at index (2,2)).