# 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
![DU1](https://i.imgur.com/Ea0Pv7b.png)
![DU2](https://i.imgur.com/CTYQIUY.png)
<!-- ![DU3](https://i.imgur.com/SFyhg3C.png) -->
![DU4](https://i.imgur.com/EbAP1kg.png)
![DU5](https://i.imgur.com/0ymbW4b.png)
<!-- ![DU6](https://i.imgur.com/96kkdYa.png) -->
![DU7](https://i.imgur.com/dOpLiNG.png)
![DU8](https://i.imgur.com/Xk4qGFD.png)
![DU9](https://i.imgur.com/5zoVCwd.png)
![DU10](https://i.imgur.com/tBo26pX.png)
![DU11](https://i.imgur.com/PU03JhR.png)
![DU12](https://i.imgur.com/8NVbV58.png)
### Train Log
![train](https://i.imgur.com/trkBnx4.png)
## 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
![DeepPaint1](https://i.imgur.com/VTzu8H2.png)
![DeepPaint2](https://i.imgur.com/wS3W8L0.png)
![DeepPaint3](https://i.imgur.com/xNMnMdl.png)
![DeepPaint4](https://i.imgur.com/Bq0fr1X.png)
![DeepPaint5](https://i.imgur.com/XmvYmLn.png)
![DeepPaint6](https://i.imgur.com/FATdxLE.png)
## 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
![style2paint1](https://i.imgur.com/VbIceI6.png)
![style2paint2](https://i.imgur.com/QXgcFgC.png)
![style2paint3](https://i.imgur.com/waWRxUY.png)
![style2paint4](https://i.imgur.com/RbAafXs.png)
## 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
![vggadd1](https://i.imgur.com/l2reucJ.png)
![vggadd2](https://i.imgur.com/vUMMc2u.png)
![vggadd3](https://i.imgur.com/9t9JFTC.png)
## 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
![vggconcat1](https://i.imgur.com/OXGhdqO.png)
![vggconcat2](https://i.imgur.com/EPvoVsY.png)
![vggconcat3](https://i.imgur.com/fnZKQfE.png)
## 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
![pix2pix1](https://i.imgur.com/swgvRAl.png)
没有合适的资源?快使用搜索试试~ 我知道了~
动漫素描图像的自动着色_python_代码_下载
共35个文件
py:21个
png:8个
md:2个
1.该资源内容由用户上传,如若侵权请联系客服进行举报
2.虚拟产品一经售出概不退款(资源遇到问题,请及时私信上传者)
2.虚拟产品一经售出概不退款(资源遇到问题,请及时私信上传者)
版权申诉
5星 · 超过95%的资源 1 下载量 54 浏览量
2022-06-17
11:13:33
上传
评论 1
收藏 13MB ZIP 举报
温馨提示
使用生成对抗网络 (GAN) 的自动动漫素描着色 效果展示: https://camo.githubusercontent.com/d0ecfd623a849a4fd58a0be640936b5832ba2990639813571cdeb1ce972ef040/68747470733a2f2f692e696d6775722e636f6d2f46307a75446e592e706e67 https://camo.githubusercontent.com/22903e46a50f09bec855610dd9c01f8f386441a78cdceb36a65b2e9e52d5bb8d/68747470733a2f2f692e696d6775722e636f6d2f5a464176396c722e706e67 https://camo.githubusercontent.com/49080a866f8fc746fc914644df408437c8625192a656289e9c716124cdff3a9f/68747470733a2f2f692e696d6775722e636f6d2f6e3
资源推荐
资源详情
资源评论
收起资源包目录
AttentionedDeepPaint-master.zip (35个子文件)
AttentionedDeepaster
models
patch_gan.py 2KB
attention.py 2KB
__init__.py 150B
deepunet.py 6KB
colorgram
colorgram.py 6KB
colorize.py 3KB
train.sh 137B
data
test
test2.png 97KB
test1.png 153KB
styles
style2.png 190KB
style5.png 360KB
style1.png 270KB
style4.png 282KB
style6.png 274KB
style3.png 208KB
train.py 2KB
trainer
__init__.py 45B
trainer.py 970B
deepunet.py 11KB
results
README.md 5KB
poster.pdf 11.55MB
requirements.txt 309B
preprocess
sketch.py 1KB
dataloader.py 4KB
__init__.py 301B
image.py 4KB
.gitignore 4KB
extract_colorgram.py 2KB
README.md 3KB
utils
io.py 2KB
average.py 762B
losses.py 515B
__init__.py 212B
args.py 3KB
image.py 1KB
共 35 条
- 1
资源评论
- tony_66938162024-01-15终于找到了超赞的宝藏资源,果断冲冲冲,支持!
快撑死的鱼
- 粉丝: 1w+
- 资源: 9156
下载权益
C知道特权
VIP文章
课程特权
开通VIP
上传资源 快速赚钱
- 我的内容管理 展开
- 我的资源 快来上传第一个资源
- 我的收益 登录查看自己的收益
- 我的积分 登录查看自己的积分
- 我的C币 登录后查看C币余额
- 我的收藏
- 我的下载
- 下载帮助
安全验证
文档复制为VIP权益,开通VIP直接复制
信息提交成功