I am trying to plot some images using the code below:
import numpy as np
import matplotlib.pyplot as plt
from PIL import Image
if __name__ == "__main__":
image2 = Image.open("Test Images/i3.png")
image2 = np.asarray(image2)
plt.imshow(image2, cmap="gray")
plt.tight_layout()
plt.show()
But it plots something different from what I see in an image viewer:
Why is this happening?
The original image:
This is because that PNG has a palette, so pixel data are palette indices. np.asarray takes raw pixel data, the palette is not taken into account.
Use .get_palette() is not None on the PIL image object to detect if the image has a palette and .convert() to convert pixel data to "real" colors.
This is a part of raw data, to get you an idea what it looks like. It includes a corner of the capital 'C' in the inscription:
The strange thing here that catches the eye is that while black is 0 (expectedly), white is for some reason 1 (rather than the expected 255) and other colors are higher but still small, the highest value being 20. Which hints that this is a palette thing.
When you .imshow this data, it's normalized to [0,1], mapping those 20 points, in order, to points equally apart from each other on the spectrum of the Colormap used.
png-parser can show the palette data:
$ png-parser -d amdNt.png
<...>
Filename: amdNt.png | Size: 2925
(0)
IHDR:
Data size : 13
- Width : 225
- Height : 225
- Bit depth : 8
- Color type : Code = 3 ; Depth Allow = [1, 2, 4, 8] ; Each pixel is a palette index; a PLTE chunk must appear.
- Compression method : 0
- Filter method : 0
- Interlace method : 0
(1)
PLTE:
Data size : 69
b'\x00\x00\x00\xff\xff\xff\xfe\xfe\xfe\x01\x01\x01\xfd\xfd\xfd\xb4\xb4\xb4\xb2\xb2\xb2\xb6\xb6\xb6\xaf\xaf\xaf\x05\x05\x05\xfa\xfa\xfa\x10\x10\x10\xbb\xbb\xbb\x16\x16\x16\xb8\xb8\xb8\x0e\x0e\x0e\xf5\xf5\xf5\xaa\xaa\xaa\xc0\xc0\xc0\xf0\xf0\xf0\x19\x19\x19\xc4\xc4\xc4\xa6\xa6\xa6'
<...>
as well as the palette itself:
Have a look here in order to understand palletised (or indexed) images.
The short story is:
if you want and expect a greyscale image with PIL, then you should always use:
im = Image.open("filename").convert('L')
otherwise you may get a palette image.
if you want and expect an RGB image with PIL, then you should use:
im = Image.open("filename").convert('RGB')
If you don't do that, PIL may give you a palette image.
Furthermore, if you are opening images with PIL, it's generally a good idea to display them with PIL and save them with PIL, else you may run into issues of RGB/BGR ordering if you open with PIL and display with OpenCV, or with colormaps if you display them with matplotlib. I am not saying you cannot interoperate, I am just saying it is somewhat harder to mix libraries without issues of channel ordering and value ranges.
