date: 2023-08-15

RGBA to RGB Trouble shooting

Pillow library is not reliable.

When changing tensor to Pillow’s image, Pillow changes the range 0~1 to 0~255 automatically.

It does not apply the real tensor value.

To apply the real tensor value, we have to use OpenCV.

RGBA to RGB image

• OpenCV, Pillow library both change RGBA image to RGB by simply getting rid of alpha channel. not applying alpha channel to R,G,B channels. I think this is a bug in OpenCV, Pillow library.

• To correctly apply alpha channel to RGB channels, follow the details.

  • Normalise the RGBA values so that they’re all between 0 and 1 - just divide each value by 255 to do this. We’ll call the result Source.

  • Normalise also the matte colour (black, white whatever). We’ll call the result BGColor Note - if the background colour is also transparent, then you’ll have to recurse the process for that first (again, choosing a matte) to get the source RGB for this operation. (a black background - BGColor to 0,0,0. a white background - BGColor to 255,255,255.)

  • Now, the conversion is defined as (in complete psuedo code here!):

    Source => Target = (BGColor + Source) =
    Target.R = ((1 - Source.A) * BGColor.R) + (Source.A * Source.R)
    Target.G = ((1 - Source.A) * BGColor.G) + (Source.A * Source.G)
    Target.B = ((1 - Source.A) * BGColor.B) + (Source.A * Source.B)
    

  To get the final 0-255 values for Target you simply multiply all the normalised values back up by 255, making sure you cap at 255 if any of the combined values exceed 1.0 (this is over-exposure and there are more complex algorithms dealing with this that involve whole-image processing etc.).

• Actual Python code for changing RGBA tensor to RGB tensor

# rgba_tensor : [batch_size, channel, height, width]
rgba_tensor = torch.mul(255,torch.randn(16, 4, 64, 256)) 
rgba_tensor = torch.clamp(rgba_tensor, 0, 255)

BG_color = (255,255,255) # white background
R_bg, G_bg, B_bg = BG_color

bs, c, h, w = rgba_tensor.size()
# rgb_tensor : rgba_tensor to RGB tensor 
rgb_tensor = torch.zeros(bs, 3, h, w)

r_ch, g_ch, b_ch, a_ch = rgba_tensor[:,0,:,:], rgba_tensor[:,1,:,:], rgba_tensor[:,2,:,:], rgba_tensor[:,3,:,:]
rem_a_ch = torch.sub(1., a_ch)

rgb_tensor[:,0,:,:] = torch.add(torch.mul(r_ch, a_ch), torch.mul(R_bg, rem_a_ch))
rgb_tensor[:,1,:,:] = torch.add(torch.mul(g_ch, a_ch), torch.mul(G_bg, rem_a_ch))
rgb_tensor[:,2,:,:] = torch.add(torch.mul(b_ch, a_ch), torch.mul(B_bg, rem_a_ch))

tensor to numpy array

• opencv imwrite 하려면 BGR 순서로 바꿔줘야.
• image 로 표현하려면 범위체크 반드시! 0~1로 되어있다면 *255 필요

np_arr = tensor.detach().to('cpu').numpy().transpose(1, 2, 0)