Numpy / Scipy带有蒙版和RGB图像

Question

I'm trying to create a mask for an RGB image using skikit learn. I want to create a mask selecting only pixels which are equal to [0,10,0], ie 10 on green channel. And then show only those pixels. This should be straight-forward, akin to http://scikit-image.org/docs/dev/user_guide/numpy_images.html , but I'm struggling.

If image is a loaded jpg, I can do

mask = image == [0,10,0]
image = image[mask]
viewer = ImageViewer(image)
viewer.show()

However, I get:

TypeError: Invalid dimensions for image data

If I then print(mask), I realise that instead of a series of True and False for each pixel, I'm getting:

[[ True False  True]
  [ True False  True]
  [ True False  True]
  ..., 
  [ True False  True]
  [ True False  True]
  [ True False  True]]

Note that the 1st pixel in my image is black. So it appears what it's doing is comparing [0,0,0] with [0,10,0], and instead of raising False, it raises True,False,True.

This then appears to fail as a mask, because I have 3 masks for each pixel, I suppose!

Does anyone know a simple way to get this masking working with RGB images?

Thanks

Answer 1

You could get that 2D mask with ALL reduction along the last axis -

mask = (image == [0,10,0]).all(-1)

Then, image[mask] would be (N,3) shaped array of only [0,10,0] values, where N is number of pixels which were of that specific RGB triplet.

So, the step(s) to using mask to show the masked image or overlay would depend on the viewer.

---

For an in-situ edit in the image, such that we would mask out everything that's not of that specific RGB triplet, we could multiply with the mask -

image *= mask[...,None]

Or create a copy with a choosing mechanism using np.where -

image_overlayed = np.where(mask[...,None], image, 0)

To get a 3D mask (if that's what needed with viewer), we could replicate the mask along the channels as well -

np.repeat(mask[...,None],3,axis=2)