动漫素描图像的自动着色_python_代码_下载

# Experiment Results ## DeepResidualPaint Application ### Configuration - Based on SegUnet with Color Histogram - Replaced single convolutional layer with Residual Block - Removed Guide Decoder - Removed Feature Extraction (Just Color Histogram) Hyperparameter | Value -------------- | --------- `learning_rate` | `0.0002` `beta1 (optimizer)` | `0.5` `batch_size` | `4` `lambda` | `100` `epochs` | `15` `learning_rate schedule` | `None` `Discriminator` | `PatchGAN` ### Discussion - Let's apply attention mechanism (maybe our final work) - How about extract color histogram for 4 seperated region (top to bottom) ### Examples Sketch - Style - Output - Ground Truth - Color Histogram of Style   <!--  -->   <!--  -->       ### Train Log  ## DeepPaint Application ### Configuration - Our idea (Using SegUnet with Colorgram) - Use VGGNet19_bn to extract feature - Used 2 guide decoder - Train Setting Hyperparameter | Value -------------- | --------- `learning_rate` | `0.0002` `beta1 (optimizer)` | `0.5` `alpha (loss)` | `0.3` `beta (loss)` | `0.9` `batch_size` | `4` `lambda` | `150` `epochs` | `200` `learning_rate schedule` | `None` ### Problem - It converts style much better than previous models - But still cannot get style for dark image - Cannot colorize well for noisy background image - How to notice portion of color? - How about using attention mechanism - More deep generator and more deep discriminator - Using different feature extraction model (for tag prediction model) ### Examples Sketch - Style - Guide1 - Guide2 - Output - Ground Truth - Colorgram of Style       ## Style2Paint Application ### Configuration - Using Style2Paint Original Paper (But using patch-gan discriminator) - Add 2 guide decoder - Train Setting Hyperparameter | Value -------------- | --------- `learning_rate` | `0.0002` `beta1 (optimizer)` | `0.5` `alpha (loss)` | `0.3` `beta (loss)` | `0.9` `batch_size` | `4` `lambda` | `50` `epochs` | `200` `learning_rate schedule` | `None` ### Problem - Still, cannot import style reference well... - Noisy background appear - How to solve :( ### Examples Sketch - Style - Guide1 - Guide2 - Output - Ground Truth     ## VGGUNet-add Application ### Configuration - Using VGGNet19_bn as feature extractor as style2paint paper - Add 2 guide decoder - Use ground truth as style hint in training and use arbitrary style in inference - add extracted feature (not concat) - Train Setting Hyperparameter | Value -------------- | --------- `learning_rate` | `0.0002` `beta1 (optimizer)` | `0.5` `alpha (loss)` | `0.3` `beta (loss)` | `0.9` `batch_size` | `1` `lambda` | `10` `epochs` | `200` `learning_rate schedule` | `None` ### Problem - Colorize well, but cannot apply hint well yet... ### Examples Sketch - Style - Output - Ground Truth    ## VGGUNet-concat Application ### Configuration - Using VGGNet19_bn as feature extractor as style2paint paper - Add 2 guide decoder - Use ground truth as style hint - concat extracted feature (not add) - Train Setting Hyperparameter | Value -------------- | --------- `learning_rate` | `0.0002` `beta1 (optimizer)` | `0.5` `alpha (loss)` | `0.3` `beta (loss)` | `0.9` `batch_size` | `4` `lambda` | `10` `epochs` | `200` `learning_rate schedule` | `None` ### Problem - Cannot use arbitrary hint ### Examples Sketch - Ground Truth - Output    ## Pix2Pix Application ### Configuration - Using original pix2pix architecture. - Simple modification to meet 512x512 resolution. - Train Setting Hyperparameter | Value -------------- | --------- `learning_rate` | `0.0002` `beta1` | `0.5` `batch_size` | `4` `lambda` | `100` `epochs` | `200` `learning_rate schedule` | `None` ### Problem - Do not use style hint - Biased pattern appeared ### Examples Sketch - Ground Truth - Output