Why does ISWT2 (R2017a) fail for coefficients in single precision?

ISWT2 (R2017a) seems not to work properly for coefficients in single precision. The code below fails only for the single precision case.

nLevels = 1;

%%For image in double
X = im2double(imread('cameraman.tif'));
C = swt2(X,nLevels,'haar');
Y = iswt2(C,'haar');
psnr(Y,X)

%%For image in single
X = im2single(imread('cameraman.tif'));
C = swt2(X,nLevels,'haar');
Y = iswt2(C,'haar'); % <- error
psnr(Y,X)

Why does this phenomenon happen?

Answer 1

[Walter Roberson]

It is a bug in the code. If you look near line 110 of iswt2 you will see

    % Ambiguity:
    % (level=3 and indexed BW image) or (level=1 and truecolor image)
    % To suppress this Ambiguity, the function SWT2 in case of a true 
    % color image and a level 1 analysis, produce single for approximation
    % coefficients !!

So the swt2 code is encoding information about BW vs truecolor in the datatype (!) and then iswt2 flubs the reconstruction.

The work-around is to use

[A,H,V,D] = swt2(X,nLevels,'haar');
Y = iswt2(A, H, V, D);

[Walter Roberson]

I just wrote this up as a bug report.

[Shogo Muramatsu]

I understood the cause clearly. Thank you very much!

[Walter Roberson]

The problem occurs for any grayscale data that is not of type double

Answer 2

[Ankit Bhatnagar]

Hi,

The ISWT2 works only with double precision as it uses functionality which expects the value to be double. Kindly use the double precision for this function.

[Walter Roberson]

No, it's just a bug. Have a look at my Answer.