It's magic. R2017a gives a different result from R2016a/b with the following simple code...

Question

Shogo Muramatsu on 10 Apr 2017

1
Link

Direct link to this question

https://nl.mathworks.com/matlabcentral/answers/334540-it-s-magic-r2017a-gives-a-different-result-from-r2016a-b-with-the-following-simple-code

Answered: Prerana Paravastu on 20 Nov 2017

Accepted Answer: Walter Roberson

Open in MATLAB Online

R2017a gives a different result from R2016a/b with the following code:

I = magic(16); H = magic(4); imfilter(I,H,'circular')

Has the implementation of imfilter been changed?

It is a huge problem for me.

4 Comments
Show 2 older commentsHide 2 older comments

Adam on 11 Apr 2017

I'm not sure if a bug being improved is a good thing or a bad thing!

Shogo Muramatsu on 11 Apr 2017

Edited: Shogo Muramatsu on 11 Apr 2017

This bug is a fatal problem especially for researchers and developers concerning signal processing, and should be fixed immediately.

For a double precision input image, the circular extension may fail when using an even size kernel. Please also see the answer made by myself.

Circular convolution is a special operator because

It is diagonalized by the discrete Fourier transform(DFT), i.e., FFT.
The adjoint operator is also circular convolution with flipped kernel, i.e., circular correlation.

This bug is an obstacle to using these properties on MATLAB.

Sign in to comment.

Sign in to answer this question.

Answer 1

Walter Roberson on 10 Apr 2017

1
Link

Direct link to this answer

https://nl.mathworks.com/matlabcentral/answers/334540-it-s-magic-r2017a-gives-a-different-result-from-r2016a-b-with-the-following-simple-code#answer_262433

Yes, the implementation did change; I do not see any release notes saying so, though.

I see in R2016b that non-separable filters have two branches: filterSingle2DWithConv() for single(), and filterPartOrWhole() for everything else. filterPartOrWhole() calls mex routines to do the work. You are using double(), so you would get the "everything else" mex routines.

I see in R2017a that non-separable filters have three branches: filterDouble2DWithConv() for double() for 2D arrays; filterDouble2DWithConv() (same routine) for double() with 3D arrays; and filterPartOrWhole() otherwise. (Notice filterSingle2DWithConv is gone and handled by filterPartOrWhole() now.) You are using double(), so you get the new filterDouble2DWithConv() routine. It calls upon conv2() to do the work.

I think you would be justified in filing a bug report on this.

4 Comments
Show 2 older commentsHide 2 older comments

Shogo Muramatsu on 10 Apr 2017

R2016a/b give correct results for the above commands.

Shogo Muramatsu on 11 Apr 2017

Open in MATLAB Online

I've confirmed that the single precision routine gives the correct result by the following code:

% Input 2D array
I = magic(16);
% Filter kernel
H = magic(4);
% Circular convolution with double precision
ansimfd = imfilter(double(I),H,'conv','circular');
% Circular convolution with single precision
ansimfs = imfilter(single(I),H,'conv','circular');
% Frequency domain filtering
Iw = fft2(I,16,16);
Hw = fft2(H,16,16);
ansfft2 = circshift(ifft2(Iw.*Hw,16,16),-[2 2]);
% Evaluation of double precision filtering
norm(ansimfd-ansfft2)
% Evaluation of single precision filtering
norm(ansimfs-ansfft2)

Casting the input array to single precision would be a way to avoid the malfunction for the double precision data.

Sign in to comment.

Answer 2

Shogo Muramatsu on 11 Apr 2017

0
Link

Direct link to this answer

https://nl.mathworks.com/matlabcentral/answers/334540-it-s-magic-r2017a-gives-a-different-result-from-r2016a-b-with-the-following-simple-code#answer_262643

Edited: Shogo Muramatsu on 12 Apr 2017

Open in MATLAB Online

From line 450 to line 451 in imfilter.m, we can see the following codes:

 a = padarray_algo(a, prePadSize, method, padVal, 'pre');
 a = padarray_algo(a, padSize, method, padVal, 'post');

These two lines seem to periodically expand the image boundary. In this second line, it seems that the periodicity is collapsing as a result of postpadding by duplicating the result of prepadding.