Matlab crash after using 'noise' (poisson) on a simple test image

[tmtenbrink]

Component
DIPimage (DIPlib 3.0.0 release, https://github.com/DIPlib/diplib/releases/tag/3.0.0), compiled from source

Describe the bug
I am testing some Matlab code for Fourier Ring Correlation, but when attempting to use the noise function with 'poisson' as noise_type on a simple test image, Matlab itself crashes. Guassian does seem to work, so it seems to be a specific problem with poisson. When the function is executed, Matlab freezes and a Microsoft Visual C++ Runtime Library error window pops up, describing some additional details. It seems like an invalid mean is passed to the random poisson function. See the image below.

The buttons are not clickable as Matlab crashes and exits after a few seconds.

To Reproduce

1. Create a simple test image a = (yy('corner'))*sin((xx('corner'))^2/300);
2. Run the poisson noise function on it a1 = noise(a,'poisson',1); %shotnoise 1

I also tested this by importing a TIFF image, which fails in the same way.

System information:

• What OS do you have? Windows 10
• If DIPimage, what version of MATLAB do you have? MATLAB Version: 9.10.0.1739362 (R2021a) Update 5
• If compiled from sources, what compiler and version did you use? I used CMake 3.21.2 with Microsoft (R) C/C++ Optimizing Compiler Version 19.28.29915 for x64 (Microsoft Visual Studio 16 2019)

[crisluengo]

Poisson noise is added by generating, for each pixel, a Poisson-distributed random value with that pixel's value as mean. A Poisson random value must have a positive mean.

The C++ Poisson generator that comes with my compiler works just fine with non-positive mean, generating 0 as output. But there is an assert in some C++ standard libraries that verifies that the mean is not <=0. When building DIPlib/DIPimage in a release build, the assert is ignored, and things work just fine. It is explicitly the debug build that will error out this way. I have fixed DIPlib to call the C++ Poisson generator only if the input pixel is > 0.

Thanks for your report!

[tmtenbrink]

Ah, I was unaware the pixels could have negative values, I had assumed they would be nonnegative. Thank you for the information!

One question: you say this should not happen in a release build, but isn't the 3.0.0 version a release build? Or must I have compiled it incorrectly from source?

[crisluengo]

@tmtenbrink You triggered an assert, which are typically disabled in release builds, and active only in debug builds. In MSVC, ensure that you select "release" for build mode. I have very little experience with MSVC (or anything Windows), so I don't remember exactly what these build modes are called there, or what the default build mode is. But I'm fairly sure you build in debug mode.

In C++ there is an assert statement, which is turned on in debug mode and turned off in release mode. There is also the DIP_ASSERT macro, which we use in the DIPlib and DIPimage code. This one is turned on or off using a setting in CMake. It was an assert statement that was triggered, not a DIP_ASSERT macro.