How can I efficently pad an RGB numpy array with the median of the image?

Question

I have RGB images which have already been rescaled so that the longer edge becomes 256 pixels, now I want to pad the border with the median RGB values of that image so the resulting image is always 256×256 pixels.

This code already works, but I am sure there could be a simpler more elegant way to do this:

img = loadAndFitImage(filePath, maxSideLength=256, upscale=True)
shp = img.shape
#the shp in this case is typically (256,123,3) or (99,256,3)

leftPad = (256 - shp[0]) / 2
rightPad = 256 - shp[0] - leftPad
topPad = (256 - shp[1]) / 2
bottomPad = 256 - shp[1] - topPad

# this part looks like there might be a way to do it with one median call instead of 3:
median = (np.median(img[:, :, 0]),np.median(img[:, :, 1]),np.median(img[:, :, 2]))

img = np.lib.pad(img, ((leftPad,rightPad),(topPad,bottomPad),(0,0)),
 'constant',constant_values=0)

if leftPad > 0:
    img[:leftPad,:,0].fill(median[0])
    img[:leftPad,:,1].fill(median[1])
    img[:leftPad,:,2].fill(median[2])
if rightPad > 0:
    img[-rightPad:,:,0].fill(median[0])
    img[-rightPad:,:,1].fill(median[1])
    img[-rightPad:,:,2].fill(median[2])
if topPad > 0:
    img[:,:topPad,0].fill(median[0])
    img[:,:topPad,1].fill(median[1])
    img[:,:topPad,2].fill(median[2])
if bottomPad > 0:
    img[:,-bottomPad:,0].fill(median[0])
    img[:,-bottomPad:,1].fill(median[1])
    img[:,-bottomPad:,2].fill(median[2])

Edit (Additional Info):

This is how the final result should look like:
This is how it should not look like (median per column):

Asked By: Quasimondo

||

Source

Answer 1

Calculating the median can be done with median = np.median(img.reshape(-1, 3), axis=0) or something similar, see this answer.

The padding can probably be done with a single line per side, something like img[:leftPad,:,:] = median. Have a look at the broadcasting rules.

Answered By: Bas Swinckels

Answer 2

You can do it easily with:

import numpy as np

a = np.asarray([[1,2,3,4,5,6],
     [8,4,5,6,7,7],
     [1,2,3,4,5,6],
     [1,2,3,4,5,6],
     [1,2,3,4,5,6],
     [1,2,3,4,5,6]])

b = a * 3
c = a * 4
d = (a,b,c)

im = np.asarray([np.pad(x, (2,), 'constant', constant_values=(np.median(x) ,)) for x in d])
print im

Output:

[[[ 4  4  4  4  4  4  4  4  4  4]
  [ 4  4  4  4  4  4  4  4  4  4]
  [ 4  4  1  2  3  4  5  6  4  4]
  [ 4  4  8  4  5  6  7  7  4  4]
  [ 4  4  1  2  3  4  5  6  4  4]
  [ 4  4  1  2  3  4  5  6  4  4]
  [ 4  4  1  2  3  4  5  6  4  4]
  [ 4  4  1  2  3  4  5  6  4  4]
  [ 4  4  4  4  4  4  4  4  4  4]
  [ 4  4  4  4  4  4  4  4  4  4]]

 [[12 12 12 12 12 12 12 12 12 12]
  [12 12 12 12 12 12 12 12 12 12]
  [12 12  3  6  9 12 15 18 12 12]
  [12 12 24 12 15 18 21 21 12 12]
  [12 12  3  6  9 12 15 18 12 12]
  [12 12  3  6  9 12 15 18 12 12]
  [12 12  3  6  9 12 15 18 12 12]
  [12 12  3  6  9 12 15 18 12 12]
  [12 12 12 12 12 12 12 12 12 12]
  [12 12 12 12 12 12 12 12 12 12]]

 [[16 16 16 16 16 16 16 16 16 16]
  [16 16 16 16 16 16 16 16 16 16]
  [16 16  4  8 12 16 20 24 16 16]
  [16 16 32 16 20 24 28 28 16 16]
  [16 16  4  8 12 16 20 24 16 16]
  [16 16  4  8 12 16 20 24 16 16]
  [16 16  4  8 12 16 20 24 16 16]
  [16 16  4  8 12 16 20 24 16 16]
  [16 16 16 16 16 16 16 16 16 16]
  [16 16 16 16 16 16 16 16 16 16]]]

Or in you particular case:

import numpy as np
from PIL import Image

filePath = '/home/george/Desktop/img.jpg'

Img = Image.open(filePath)
img = np.asarray(Img, np.int)
shp = np.shape(img)
img = img.transpose(2,0,1).reshape(3,215,215)

leftPad = round(float((255 - shp[0])) / 2)
rightPad = round(float(255 - shp[0]) - leftPad)
topPad = round(float((255 - shp[1])) / 2)
bottomPad = round(float(255 - shp[1]) - topPad)
pads = ((leftPad,rightPad),(topPad,bottomPad))

img_arr = np.ndarray((3,255,255),np.int)
for i,x in enumerate(img):
    cons = np.int(np.median(x))
    x_p = np.pad(x,pads,
                'constant', 
                 constant_values=cons)
    img_arr[i,:,:] = x_p

im_shp = np.shape(img_arr)
ii = np.uint8(img_arr).transpose(1,2,0)

im = Image.fromarray(np.array( (ii) ))
im.show()
im.save((filePath), "JPEG")

Output:

Answered By: Geeocode

Answer 3

I was also struggling on this and figured out an elegant answer:

color = np.median(img, axis=(0,1)) img = np.stack([np.pad(img[:,:,c], pad, mode='constant', constant_values=color[c]) for c in range(3)], axis=2)

Answered By: 黄梁华

Answer 4

Late to the party but here goes another suggestion:

def pad_image(img, color, border_width=.1):
    """
    pads image img with given color. 
    Color must be in same color space as image (usually, RGB). 
    border_width is expected to be the fraction of padding you want to add,
    with respect to the shorter dimension of the image.
    """
    h, w, c = img.shape

    # compute the number of pixels you'll pad
    border = int(float(min(h, w) * border_width))  

    # compute the "new background"
    result = np.full((h+2*border, w+2*border, c), color, dtype=img.dtype)
    
    # now fill this "new background" with your original image in the center
    result[border:-border, border:-border] = img

    return result

Usage example:

img = np.zeros((11,21, 3), dtype=np.uint8)  # create simple black image

img = pad_image(img, (128,0,0), .25)

plt.imshow(img)

Which outputs this:

For this particular case, as other said, you need to compute the median color before padding:

color = np.median(img, axis=(0,1))

Answered By: Rodrigo Laguna

Answer 5

I believe that this can be done more simply using a single np.pad() call:

  value = np.median(image, axis=(0, 1))
  pad = 2
  cval = np.array([[value, value], [value, value], [0, 0]], dtype=object)  # Ragged.
  image2 = np.pad(image, ((pad, pad), (pad, pad), (0, 0)), constant_values=cval)

Answered By: Hugues

How can I efficently pad an RGB numpy array with the median of the image?

Question:

Answers: