Get unique pixels in image

Question

The following:

image = cv2.imread("image.png")
print(image.shape)
print(np.unique(image,axis=1))

if giving:

(700, 500, 3)
[[[255 255 255]
  [255 255 255]
  [255 255 255]
  ...
  [255 255 255]
  [255 255 255]
  [255 255 255]]

 [[255 255 255]
  [255 255 255]
  [255 255 255]
  ...
  [255 255 255]
  [255 255 255]
  [255 255 255]]

 [[255 255 255]
  [255 255 255]
  [255 255 255]
  ...
  [255 255 255]
  [255 255 255]
  [255 255 255]]

 ...

 [[255 255 255]
  [255 255 255]
  [255 255 255]
  ...
  [255 255 255]
  [255 255 255]
  [255 255 255]]

 [[255 255 255]
  [255 255 255]
  [255 255 255]
  ...
  [255 255 255]
  [255 255 255]
  [255 255 255]]

 [[255 255 255]
  [255 255 255]
  [255 255 255]
  ...
  [255 255 255]
  [255 255 255]
  [255 255 255]]]

How do I get the unique pixels? For example, [255 255 255] should be one of the unique pixel values.

Reshape it to 2d array with 3 columns, so that each row represents one pixel, and the columns are the color components. Then get unique along axis 0 (i.e. unique rows). -- `np.unique(image.reshape(-1,3), axis=0)` — Dan Mašek, Jan 14 '22 at 17:25
As @DanMašek suggests is good - though there is a faster way here... https://stackoverflow.com/a/59671950/2836621 — Mark Setchell, Jan 14 '22 at 18:16
@MarkSetchell doesn't that counts the number of unique pixel only, not listing them? — Quang Hoang, Jan 14 '22 at 18:57
@QuangHoang Yes, you're right - I must pay closer attention. I guess they could be un-mangled, but it's probably not worth the effort here. — Mark Setchell, Jan 14 '22 at 19:07

score 2 · Accepted Answer · answered Jan 14 '22 at 18:52

As per comments, one of the ways to do that is np.unique(image.reshape(-1,3), axis=0) but this is can be improved significantly using dimensionality reduction with numba

import numpy as np
from numba import njit, prange

@njit(parallel=True)
def _numba_dot(arr, dimshape, len_arr, len_dimshape, su):
    for i in prange(len_arr):
        for j in range(len_dimshape):
            su[i] = su[i] + arr[i][j] * dimshape[j]
        
def numba_dimreduce(arr, dtype=int):
    dimshape = np.array([1, 256, 65536])
    su = np.zeros(len(arr), dtype=dtype)
    _numba_dot(arr, dimshape, len(arr), len(dimshape), su)
    return su

>>> image = np.array([[[255, 255, 255], [255, 254, 254]], [[255, 254, 254], [254, 254, 254]]])
>>> img = image.reshape(-1, 3))
>>> img
array([[255, 255, 255],
   [255, 254, 254],
   [255, 254, 254],
   [254, 254, 254]])
>>> u, idx = np.unique(numba_dimreduce(img), return_index=True)
>>> img[idx]
array([[254, 254, 254],
       [255, 254, 254],
       [255, 255, 255]])

In case you have large images, you might like to try np.bincount-like methods for further speed ups.

Get unique pixels in image

1 Answers1