The output of the DCE-net?

[nile649]

The image of the architecture is pretty confusing since the output of the DCE-net apparently has 8 output according to the figure.2 and the paper, which doesn't make sense by looking at the equation provided.

LE(I(x); α) = I(x) + αI(x)(1 − I(x))

here, in the above equation what exactly is the output of the DCE-net.

and here what is the output

LEn(x) = LEn−1(x) + αnLEn−1(x)(1 − LEn−1(x))

[kkedich]

DCE-net will output one alpha/curve parameter map for each channel (RGB) for each iteration. Since the paper use 8 iterations as its default value, DCE-net will output 8 curve parameter maps for each channel, with a total of 24 maps.

The first equation considers alpha for higher-order curves, where α ∈ [−1, 1]. So, the output of DCE-net are 24 scalars. But, this will only produce a global enhancement without considering particular adjustments for some regions of the image. That is the intention of the pixel-wise curve, i.e., your second equation. In this case, DCE-net will output 24 curve parameter maps with the same size as your input image (input = [256, 256, 3], curve map [256, 256, 3]).

[Li-Chongyi]

Thanks @kkedich. Your explains are exactly. We will release the code soon. It will help understanding how to output curve parameter maps.