Comparing pixels against RGB value in NumPy

Question

Assume I created an image in NumPy:

image = imread(...)

And the image is RGB:

assert len(image.shape) == 3 and image.shape[2] == 3

I want to check (i.e. get a boolean mask) which pixels are (R=255, G=127, B=63) in a cleaner and efficient way:

mask = (img[:, :, 0] == 255) & (img[:, :, 1] == 127) & (img[:, :, 2] == 63)

This code worked for me. However - please correct me if I'm wrong - this code is creating three intermediate masks and, since I'm processing images, those masks will be pretty big.

I need help with:

• Is there a way to make boolean masks occupy 1 bit per element, instead of 1 byte per element?
• Is there a way to solve the same problem without creating the three intermediate boolean masks?

This code did not work:

mask = img == (255, 127, 63)

neither this:

mask = img[:, :, :] == (255, 127, 63)

since the comparison was performed element-wise, and the resulting values were performed element-wise, yielding a 3-dimension (w, h, 3) boolean mask.

• Jamal: Please add the language tag as well. I'm also not sure if this would be enough code for review (someone else may be able to determine that).
• Luis Masuelli: Added the language. Thanks. Anyway, the code I have doubts about, has one-liners like this. Other code is irrelevant to my problem.

Answer 1

Your original code can be rewritten as:

mask = np.all(img == (255, 127, 63), axis=-1)

It is a little cleaner, but not more efficient, as it still has to allocate a mask of the same size as the image.

• Luis Masuelli: So there's no way to reduce the memory usage?
• Akababa: you can always mix the vectorization of numpy with the branching of python to get different space-time tradeoffs, for example create one numpy row (third dimension) at a time and reduce them in python.