Correct color channels in upscale using array slicing

[wfng92]

Similar to issue #1175, REALSRGAN expects BGR as input instead of RGB. This implementation reused the same numpy array slicing code as #1178.

[psychedelicious]

Oooh, interesting. Did you notice any difference in results?

Edit: I did some testing with 3 random images. The resulting images definitely are different. In some cases, imperceptibly so, but in other cases, noticeable. Still very minor differences, but they are certainly there.

928 different pixels (the result from this PR is noticeably sharper in some areas):
Original implementation:

Corrected channels:

24 different pixels:
Original implementation:

Corrected channels:

32 different pixels:
Original implementation:

Corrected channels:

[wfng92]

Yes, the resulting images are different. Based on the code from official repositories, both GFPGAN and REAL-ESRGAN use cv2.imread to read the file, which decodes the image into a matrix with the color channels stored in the order of Blue, Green, Red and A (depends on the flag used)

# GFPGAN 
input_img = cv2.imread(img_path, cv2.IMREAD_COLOR)

# REAL-ESRGAN
img = cv2.imread(path, cv2.IMREAD_UNCHANGED)

Hence, the expected image input should be BGR instead of RGB.

[lstein]

Help me out here. This is the third PR that solves the GBR <-> RGB problem. How is this different from #1178?

EDIT: Nevermind. Just realized that this applies to ESRGAN and the previous PRs were for GFPGAN.

[lstein]

LGTM