Awesome

Typography-with-Decor

Typography with Decor: Intelligent Text Style Transfer <br> Wenjing Wang, Jiaying Liu, Shuai Yang, and Zongming Guo <br> CVPR, 2019

[Project Page], [Paper]

Update 190707: fix a bug in README.md and add a python version for glyph pre-processing.

Setup

Prerequisites

• Python 3 (Python 2 also works, but needs small revision.)

• Pytorch

• sklearn

• matplotlib

• opencv-python (3.4.2, other versions should work too)

Getting Started

• Clone this repo:

git clone https://github.com/daooshee/Typography-with-Decor.git
cd Typography-with-Decor

• Download pre-trained models from Google Drive.

  Save netG.pth to pre-trained/netG.pth

  Save netD.pth to pre-trained/netD.pth

  Save netSeg.pth to Segmentation/netSeg.pth

• Download datasets from Google Drive.

  Save and unzip Decoration.zip to Data/Decoration/. These are 4k icons collected from www.shareicon.net, used as decors.

  Save and unzip TextEffects.zip to Data/TextEffects/. It contains 60 different kinds of text effects with 52 English letters of 19 fonts, totally 59k images.

  These two parts are required by the one-shot fine-tuning process.

• Run mkdir cache, mkdir checkpoints, mkdir basal-text-effect-transfer and mkdir result.

Quick Testing

• Some examples (teasor and Fig. 14 in the paper) are provided in Input/. For decor segmentation + one-shot fine-tuning + basal text effect transfer + decor recomposition, simply run: <br>

python runtest.py

• Results can be found in result/.

• Variables such as gpu id, batch size, training iteration for one-shot fine-tuning can be setted in runtest.py.

Testing Custom Images

• For custom images, save target glyph image in Input/content/, save input style/glyph image pairs in Input/style/. The input glyph image should be named with _glyph.png.

• For pre-processing the glyph image, using the following commands with Matlab:

x = imread('input.png'); 

I = im2bw(x(:,:,1));
I2 = bwdist(I, 'euclidean');
I3 = bwdist(~I, 'euclidean');
x(:,:,3) = min(255, I2);
x(:,:,2) = min(255, I3);

imwrite(x,'result.png');

The following python function works too:

import cv2
import numpy as np
from scipy.ndimage.morphology import distance_transform_edt as bwdist

def PreProcess(img_path):
    img = cv2.imread(img_path)
    I = img[:,:,2]
    I2 = bwdist(I <= 100);
    I3 = bwdist(I > 100);

    img[:,:,0] = np.clip(I2,0,255);
    img[:,:,1] = np.clip(I3,0,255);

    return img

• Write the names of your custom images in runtest.py.

• Run python runtest.py.

Training Basal Text Effect Transfer

• If you are interested in training from the start, for training basal text effect transfer, first download datasets from Google Drive.

  Save and unzip Decoration.zip to Other/BasalTextEffectTransfer/Data/Decoration.

  Save and unzip TextEffects.zip to Other/BasalTextEffectTransfer/Data/TextEffects.

• cd Other/BasalTextEffectTransfer/.

• Run python train.py --gpu 0 . More parameters can be found by python train.py --help.

Training Decor Segmentation

• For training decor segmentation, first download datasets from Google Drive.

  Save and unzip Decoration.zip to Other/DecorSegmentation/Data/Decoration.

  Save and unzip TextEffects.zip to Other/DecorSegmentation/Data/TextEffects.

  Save and unzip Texture.zip to Other/DecorSegmentation/Data/Texture.

  Save and unzip WildDataMask.zip to Other/DecorSegmentation/Data/WildDataMask.

• cd Other/DecorSegmentation/.

• First train with with L1 loss for 30 epochs with a mini-batch size of 90:

python train_domain_adaptation.py --gpu 0 --batchSize 90 --niter 17000 --use_decay

• Then train with L1 loss and perceptual loss for 50 epochs with a mini-batch size of 15:

python train_domain_adaptation.py --gpu 0 --batchSize 15 --niter 170000 --use_decay --use_perceptual

• Finally train with full loss for 5000 iterations with a mini-batch size of 10.

python train_domain_adaptation.py --gpu 0 --batchSize 10 --niter 5000 --use_decay --use_perceptual --domain_adaptation

Citation

If you use this code for your research, please cite our paper:

@InProceedings{typography2019,
   author = {Wang, Wenjing and Liu, Jiaying and Yang, Shuai and Guo, Zongming},
   title = {Typography with Decor: Intelligent Text Style Transfer},
   booktitle = {The IEEE Conference on Computer Vision and Pattern Recognition (CVPR)},
   month = {June},
   year = {2019}
}