do anyone have idea why the two outputs below are different

in 1st code block image is loaded and PIL resize used. while in 2nd block keras load_img parameter: target_size is used. for same steps it is giving different output.

from keras.preprocessing.image import load_img 
import numpy as np
path = 'C:/Users/user/Downloads/random_colour_image.JPG' # actual snippet of image:https://wallpaperaccess.com/full/1523270.jpg
target_size = (3,3)

#code block 1
image = load_img(path)
image = image.resize(target_size)
image = np.asarray(image)
print(image)

Output 1:

[[[132  99  79]
  [146  80  68]
  [165  15  81]]

 [[116 102  94]
  [133 101  69]
  [198  28  53]]

 [[ 82 129 108]
  [119  89 112]
  [166  87  51]]]

code block 2:

image = load_img(path, target_size=target_size)
image = np.asarray(image)
print(image)

Output 2:

[[[ 48 190  88]
  [ 57 159  49]
  [145   0  77]]

 [[116  90 101]
  [ 14 133  67]
  [146  19   2]]

 [[  5 119  50]
  [129  69  97]
  [179  63   2]]]

CodePudding user response：

For getting the same output from two methods you need to consider:

1. In keras.preprocessing.image.load_img(target_size = (W,H)) you should not set resize output, you should insert your original image_size.
2. For resizing image after reading with keras.preprocessing.image.load_img, you need to convert PIL to numpy.array then to tensor then use tf.image.resize.
3. In tf.image.resize you need to set method='bicubic' and antialias = True to get result of two methods same as each other.

Example:

from keras.preprocessing.image import load_img 
import tensorflow as tf
import numpy as np

path = 'test.png' 
original_size = (32,32)                           # <- you want this
resize_size = (3,3)                               # <- you want this

#code block 1
image = load_img(path)
image = image.resize(resize_size)
image = np.asarray(image)
print(image)


#code block 2
image = load_img(path, target_size=original_size)  # <- you want this
image = tf.image.resize(                           # <- you want this
    tf.convert_to_tensor(np.asarray(image)) ,      # <- you want this
    resize_size,                                   # <- you want this
    method='bicubic',                              # <- you want this
    antialias = True                               # <- you want this
    )
image = np.asarray(image).astype('int')
print(image)

Output:

# Output 1
[[[132 129 123]
  [126 124 135]
  [121 130 118]]

 [[120 132 125]
  [128 132 117]
  [129 126 120]]

 [[125 120 133]
  [134 123 122]
  [137 117 125]]]

#Output 2
[[[132 128 122]
  [125 124 135]
  [121 130 117]]

 [[119 131 124]
  [127 131 117]
  [128 126 120]]

 [[125 119 132]
  [133 122 122]
  [137 117 124]]]

Creating a random image:

import numpy as np
from PIL import Image
imarray = np.random.rand(32,32,3) * 255
im = Image.fromarray(imarray.astype('uint8')).convert('RGB')
im.save('test.png')

CodePudding user response：

thanks I'mahdi. but target_size is also intended to load an resized image.
got the solution: Keras load_img default interpolation is NEAREST(0) and that of PIL.resize is BICUBIC(3). hence that ambiguity.

change in code block 1:

image = image.resize(target_size, resample=0)

or change in code block 2:

image = load_img(path1, target_size=target_size, interpolation='bicubic')

will solve it and produce same result.
Thanks.