How to down sample image array, without changing pixel values

Question

I have image segmentation project, and ground truth labels given as images where pixel value stands for the label. I need to resize the images and labels, while keeping the labels in the same value set.

I tried many things, All change the value set.

Lets create dummy data

from skimage.transform import rescale, resize
from scipy import ndimage
from PIL import Image
import cv2

mask = np.zeros((30,20), dtype=np.uint16)
mask[22:26,12:30]=70
mask[25:27,14:17]=30
print('original label', mask.shape, np.unique(mask))

Outputs: original label shape: (30, 20) original label values: [ 0 30 70]

I need to resize label, so the result will have only 0, 30, 70 values.

What I tried

skimage_resized = resize(mask, (mask.shape[0]//2, mask.shape[1]//2), mode='constant')
print(skimage_resized.shape, np.unique(mask_resized))

skimage_rescale = rescale(mask, 1.0/2.0, mode='constant')
print(skimage_rescale.shape, np.unique(mask_resized))

ndimage_resized = ndimage.interpolation.zoom(mask, 0.5)
print(ndimage_resized.shape, np.unique(mask_resized))


cv2_resized = cv2.resize(mask, (mask.shape[0]//2, mask.shape[1]//2),
                        interpolation=cv2.INTER_NEAREST)
print(cv2_resized.shape, np.unique(mask_resized))

mask_pil = Image.fromarray(mask, mode=None)
pil_resized = mask_pil.thumbnail((mask.shape[0]//2, mask.shape[1]//2), Image.NEAREST)
print(skimage_resized.shape, np.unique(pil_resized))

Output:

(15, 10) [ 0  5  6 28 29 30 31 61 62 65 70 71 74 75 76]
(15, 10) [ 0  5  6 28 29 30 31 61 62 65 70 71 74 75 76]
(15, 10) [ 0  5  6 28 29 30 31 61 62 65 70 71 74 75 76]
(10, 15) [ 0  5  6 28 29 30 31 61 62 65 70 71 74 75 76]
(15, 10) [None]

Answer 1

Found a solution with openCV.

import numpy as np
import cv2
resizeto = 2
small_lable = cv2.resize(mask, (mask.shape[1]//resizeto, 
                         mask.shape[0]//resizeto),
                        interpolation=cv2.INTER_NEAREST)
small_lable = (np.array(small_lable)).astype('uint8')
print(small_lable.shape, np.unique(small_lable))
plt.imshow(small_lable)

output:

(15, 10) [ 0 30 70]

Answer 2

From the docs (emphasis mine):

Note that when down-sampling an image, resize and rescale should perform Gaussian smoothing to avoid aliasing artifacts. See the anti_aliasing and anti_aliasing_sigma arguments to these functions.

Downscale serves the purpose of down-sampling an n-dimensional image by integer factors using the local mean on the elements of each block of the size factors given as a parameter to the function.

One possible workaround to your problem would be downsampling through basic slicing:

import numpy as np

dx, dy = 2, 2
mask = np.zeros((30, 20), dtype=np.uint16)
mask[22:26, 12:30] = 70
mask[25:27, 14:17] = 30

downsampled = mask[::dx, ::dy]
print(downsampled.shape, np.unique(downsampled))

The output from the snippet above is:

(15, 10) [ 0 30 70]