Image Preprocessing corrupts some greyscale images

[alexzubiaga]

While training an Inception Model with a mix of color and greyscale images I've noticed that some of the greyscale images were loosing most of the details when loading them with flow_from_directory.

The cause is the RGB conversion from Pillow used in load_img as described in this issue Pillow-2574

In my case the suggested workaround fixed the conversion.

This is the code I've used to handle the RGB conversion:

if img.mode != 'RGB':
    if img.mode == 'I':
        from PIL import ImageMath
        img = ImageMath.eval('im/256', {'im': img})
    img = img.convert('RGB')

[keck-in-space]

Super super helpful. Thank you!

[alexzubiaga]

And I thought I was the only one working with greyscale images. Glad to be of help.

[ronz0215]

Hello,
I have similar issue, I'm using mode 'I' too, however my images are saved as uint16 so scaling by 256 losses too much info. I noticed the numpy array can be a float32...has anyone tried to normalize to 16k?

Also, I'm not super familiar with Python and my understanding is you can't overload classes...so where did you make your changes? Did you make changes to the keras modules in your site-packages dir or is there some other way to make these changes without modifying the downloaded code?