Issue converting a PNG image to a JPG image using PIL

Question

I'm working with X-ray images to build up a neural network, using Pytorch. After some fumbling, I found my main issue was on initial loading of the images. For example, I start with this image:

When I try to load it using the ImageFolder, and then DataLoader, it shows as:

After looking through it, I've realised that it is PIL.show() causing the issue (and the internal PNG-JPG conversion), but for the life of me, I can't get it to work. I've tried PIL's .convert() method, with no luck, as well as the steps for copying an alpha mask onto a background as shown here, I've tried changing the background colour to black as well, and nothing, everything just degrades into the white box. I've tried it on other X-rays in the files as well, and it's the same result.

When I tried to convert an image using PNGtoJPG.com, I got this image: , but FreeConverter.com worked fine.

I know it must be something with the transparency of the original image, but at this point, I've tried everything I can find, and nothing.

For reference, this is the conversion code I've been testing on:

png = Image.open("0001_1297860395_01_WRI-L1_M014.png") #with the correct file path
    png.load()
    alpha = png.split()[-1]
    alpha= alpha.convert('RGBA')
    background = Image.new("RGB", png.size, (0, 0, 0))
    background.paste(png, alpha)
    background.save('./foo3.jpg', 'JPEG', quality=80)

The images load fine when using cv2, but since I need to use a custom transform, the method needs to convert into cv2.

Any help would be greatly appreciated, I'm at a loss!

Answer 1

The individual image conversion can be handled by using png.point(lambda x: x / 256) This is documented here: https://github.com/python-pillow/Pillow/issues/6765, where it's shown that PIL has issues with PNGs. However, the issue of it not carrying through the neural network is still an issue, but better suited for another question.

Answer 2

Your image is a 16-bit, greyscale PNG. You can see that with exiftool:

exiftool -BitDepth -ColorType image.png

Output

Bit Depth                       : 16
Color Type                      : Grayscale

It opens in I mode within PIL, i.e. as 32-bit signed integers:

from PIL import Image

im = Image.open('image.png')
print(im)
print(im.mode)

That prints:

<PIL.PngImagePlugin.PngImageFile image mode=I size=494x727>
I

An easy way to make it 8-bit unsigned is convert to Numpy array, shift right 8 bits and convert to unsigned 8-bit:

# Make "im" into Numpy array, shift right 8 bits, convert to uint8
na = (np.array(im)>>8).astype(np.uint8)

Now you can make that back into PIL Image with:

pilImage = Image.fromarray(na)
pilImage.show()