Filter Numpy array based on different size mask array

Question

I'm trying to mark regions of an image array (224x224) to be ignored based on the value of a segmentation network's class mask (16x16). Previous processing means that unwanted regions will be labelled as -1. I want to be able to set the value of all regions of the image where the class mask reports -1 to some nonsense value (say, 999), but maintain the shape of the array (224x224). A concise example of what I mean is below, using a 4x4 image and 2x2 mask.

# prefilter
image = 1 2 3 4
        5 6 7 8
        9 1 2 3
        4 5 6 7

mask  = -1 4
        -1 5

# postfilter

image_filtered = 999 999 3 4
                 999 999 7 8
                 999 999 2 3
                 999 999 6 7

Is there an efficient way to do this modification?

Will the mask always be an integer multiple of the image? ie. image = 14x14, mask =7x7, not image =15x15, mask =6x6 — Lone Survivr
– Lone Survivr, Commented Jan 13, 2022 at 17:33
@LoneSurvivr Yes. I can control the size of the mask, but want to keep it roughly close to the network output (13x16), to reduce the possible error from upscaling. — sarobando
– sarobando, Commented Jan 13, 2022 at 17:43
Is the area masked always going to be a rectangle, or will it be freeform? — Lone Survivr
– Lone Survivr, Commented Jan 13, 2022 at 20:01

Lone Survivr · Accepted Answer · 2022-01-14 01:49:58Z

3

Here's some code I got to work. It does require the mask and the image to have the same aspect ratio, and be integer multiples of each others sizes.

import numpy as np

image = np.array([[1,2,3,4],
                  [5,6,7,8],
                  [9,1,2,3],
                  [4,5,6,7]])

mask = np.array([[-1,4],
                 [-1,5]])

def mask_image(image, mask, masked_value):
    
    scale_factor = image.shape[0]//mask.shape[0] # how much the mask needs to be scaled up by to match the image's size
    
    resized_mask = np.kron(mask, np.ones((scale_factor,scale_factor))) # resize the mask (magic)

    return(np.where((resized_mask==-1),masked_value,image)) # where the mask==-1, return the masked value, else return the value of the original array.
    
print(mask_image(image, mask, 999))

I used np.kron to resize an array after seeing this answer.

This function is ~~extremely fast, it took ~2 sec to mask a 1920x1080 image with a 1920x1080 mask~~.

EDIT: Use what @Jérôme Richard said in their comment.

edited Jan 14, 2022 at 1:49

answered Jan 13, 2022 at 20:25

Lone Survivr

3721 silver badge9 bronze badges

Sign up to request clarification or add additional context in comments.

2 Comments

Jérôme Richard Over a year ago

~2 sec is not really "extremely fast". Did you mean something else? On my machine it takes 45 ms in this case which is quite good. Consider using resized_mask = mask.repeat(scale_factor,axis=0).repeat(scale_factor,axis=1) instead of a cryptic slow Kronecker product. This is about 3 times faster on the same 1920x1080 use-case (15 ms).

Lone Survivr Over a year ago

You're right, I edited my answer

Collectives™ on Stack Overflow

Filter Numpy array based on different size mask array

1 Answer 1

2 Comments

Your Answer

Linked

Hot Network Questions

Collectives™ on Stack Overflow

1 Answer 1

2 Comments

Your Answer

Sign up or log in

Post as a guest

Linked

Related