Masked image translation registration

[LaurentRDC]

I'd like to gauge interest in masked image translation registration.

I often have to register translation between diffraction patterns. However, there are static image elements that throw off standard approaches based on normalized cross-correlation, like skimage.feature.register_translation.

I have implemented a normalized cross-correlation function that can take image masks as well, based on the following publication: D. Padfield. Masked object registration in the Fourier domain. IEEE Transactions on Image Processing (2012). The working principle is essentially the same as skimage.feature.register_translation, with the cross-correlation not taking into account the masked parts of images.

You can find a successful registration example here.

I have already written a Python implementation that mimics skimage.feature.register_translation as much as possible:

• The source code for the masked, normalized cross-correlation is visible here as mnxc2
• The source code for the overall masked image translation registration is visible here as masked_register_translation

The function masked_register_translation above passes the exact same tests as the MATLAB implementation from the original author. It also gives the same results as skimage.feature.register_translation in the case of trivial masks.

I'd like to get some feedback regarding including this functionality directly in Scikit-image. The source code linked above provides a working starting point.

[jni]

@LaurentRDC cool package! You should present it at a SciPy conference!

I would support inclusion of this if:

1. You could make it n-dimensional (seems doable from a cursory look at the code)
2. You could include it with minimal intrusion to both the API (adding src_mask and target_mask keyword arguments at the end) and the code (not too many additional branching if/elses inside the code). (Again, I think this seems doable.)

What do you think?

[jni]

Oh, another point: at scikit-image we use True values in the mask to point to valid locations. I think that's the opposite of what your API does? Again, trivial fix.

[LaurentRDC]

@jni Thank you for the feedback!

I understand the need for n-dimensional operations and consistent mask definition. However, it seems you are advocating for including the masked image registration functionality inside of register_translation.

How would you feed about having a separate function? The masked normalized cross-correlation requires 6 FFTs and 6 IFFTs, while normalized cross-correlation requires 2 FFTs and 1 IFFTs. The performance difference is very noticeable, especially in my use case. I typically want to register ~3000 images to a reference, sometimes with masks and sometimes without.

[jni]

@LaurentRDC note that this is an API question, not an implementation question (though I would like to see as much code sharing as possible). But in the worst case you can do:

def register_translation(src, target, ..., src_mask=None, target_mask=None):
    if src_mask is None and target_mask is None):
        result = register_translation_classic(src, target, ...)
    else:
        result = register_translation_masked(src, target, ..., 
                                             src_mask=src_mask, target_mask=target_mask)
    return result

[LaurentRDC]

Since #3334 has been merged, I'm closing this.