Masking

[tjgalvin]

As separately discussed, this is to fix the same set of pixels being fit be different components across different islands, leading to the doubling up of sources. This is a WIP and should be discussed a little further.

[coveralls]

Coverage increased (+0.02%) to 82.501% when pulling a57277b on tjgalvin:masking into 21c6b0a on PaulHancock:dev-aegean.

[tjgalvin]

I also noticed that instances where diagonal pixels were not being grouped together into a single island, and instead two separate islands were being created for later fitting. This was happening a lot for the lower signal to noise islands, where two diagonally adjacent pixels were being identified in the flood fill procedure. The default behavior of the scipy.ndimage.label function is to not connect these pixels in an island. I have added a structure argument to fix this.

from scipy.ndimage import label
import numpy as np
import matplotlib.pyplot as plt

a = np.zeros((4,4), dtype=bool)
a[1,2] = 1 
a[2,1] = 1 

l, n = label(a)
ll, nn = label(a, structure=np.ones((3,3)))

fig, (ax1, ax2, ax3) = plt.subplots(1,3, figsize=(10, 3))

ax1.imshow(
    a
)
ax1.set_title('Input image')

ax2.imshow(
    l
)
ax2.set_title('Label Islands (default)')

ax3.imshow(
    ll
)
ax3.set_title('Label Islands (structure)')

[PaulHancock]

I'm now wondering if this is related to #136 / #114

[tjgalvin]

I don't think it is. From my reading of that pull request, it is concerned with ensuring a region file is considered when doing the initial island characterisation. Although the current flood fill may be ignoring the --region option (there are some differences between the method and the newer function), I don't think it is related to these duplicate island issues.