lower intensity pixel lines after stitching

[josenimo]

Dear ashlar devs,

After visually inspecting out stitched images, I have noticed some dark pixel lines showing up in our images.
We used @jmuhlich rotation-correction branch, because we noticed large rotation issues (on that problem it works really well!), and we are unsure if this has anything to do with it.

Image 1: First nuclear stain, white arrow show these lines.

Image 2: Second nuclear stain, same location as Image1. Line only appears on left side.

Image 3: Third nuclear stain, no lines visible

Image 4: Remaining stains at the same location. In one of them the line ends there.

[jmuhlich]

Yes, I think I know why that happens. The code that "pastes" all the tiles into the final image performs cross-blending at the tile edges to average out any remaining minor image differences. When the tiles are rotated (and especially when --barrel-correction is used), my simplistic cross-blending technique can leave those small dark strips at the tile borders. I just haven't had time to confirm this hypothesis and work out a fix.

[jmuhlich]

The code in question is here:

ashlar/ashlar/utils.py

Lines 197 to 216 in eae4ae6

	def pastefunc_blend(target, img):
	"""Linear blend based on distance to unfilled space in target."""
	# This should catch actual holes but not the actual unfilled space.
	# FIXME Should generate mask from tile boundaries instead.
	hole_threshold = np.mean(target.shape)
	mask = skimage.morphology.remove_small_holes(target != 0, hole_threshold)
	dist = scipy.ndimage.distance_transform_cdt(mask)
	dmax = dist.max()
	if dmax == 0:
	alpha = 0
	else:
	alpha = dist / dmax
	# Keep target pixel values where img has value 0 (with a 1-pixel
	# dilation to clean up the edge). This is a temporary hack to support
	# image corrections that leave regions of zero pixels around the edge of
	# img, such as barrel correction and rotation.
	# FIXME Should compute the geometry of the source image mask more
	# deliberately and precisely.
	alpha[skimage.morphology.binary_dilation(img == 0)] = 1
	return target * alpha + img * (1 - alpha)

[jmuhlich]

As stated in the FIXME comments, I think this function could do a better job determining where the cross-blending should start and stop. One other issue is that the outcome is slightly dependent on the order in which the tiles are pasted in, which feels bad to me.

[josenimo]

Hey @jmuhlich ,
We checked two of our images and indeed it seems to be present for both (where we used the barrel correction). I see what the problem is. It would be awesome if this could be fixed :), but not really sure I can help. If you need the images or anything else we can help with let us know.