Closed Tedko closed 7 years ago
@taizan also, for your training input, did you eliminate all black/white manga, etc?
Im using YUV for inner calculation, because loss calculation in HSV is a little bit tricky. I removed pure B/W image from training data target but still some less colored items may are in the datasets.
Have you ever tried HSL? The loss calculation of L is easier. Though YUV should be enough for now. I am still considering how to get rid of the 'brightness' of the output.
oh, sorry for late reply, I m not tryed HSL yet.
Credit to @dsh0416 @fpsluozi
HSV/HSL is more suitable here for the task. Especially if we consider that the output image is relatively 'bright' than normal ones(input/test set). This might caused by that moving RGB simultaneously is more difficult than moving L/V in HSV/HSL.
Since RGB -> HSV/HSL is non-linear, here using RGB for training the network will consume several extra layers for doing the same thing but with less efficiency.