/TG-Net

Reconstruct visual wood texture with semantic attention. Image Inpainting.

Primary LanguagePython

TG-Net

Reconstruct visual wood texture with semantic attention. Image Inpainting. 📃Now my article has provided 50 days of free visiting.Welcome to click the link to view our work. paper

image

Requirements

  • Python3.7
  • Pytorch1.3.1
  • tqdm
  • PIL

Usage

  • You can find more information on ./configs and change the relevant training parameters in the config.yaml
  • batch_size
  • image_shape
  • mask_shape
  • lr, beta1, beta2 are the parameters of the Adam optimiser, for more information you can consult the manual on the torch website(https://pytorch.org/docs/1.1.0/_modules/torch/optim/adam.html).

Train

  • We provide WoodDataset for you to train your network. Download on Mendeley Data or Baidu Netdisk：(https://pan.baidu.com/s/1MO_iG0YpG9ZjjovTA5lwgA 提取码：1219). You can also use a face dataset or street view or your own dataset to train the model. Unfortunately, our trained models larger than 100M cannot be uploaded to git.
  • set images under ./img and mask image is placed under the main directory of the folder. THen,

python train.py

Test

image

  • You can retrain your own model using your dataset.
  • Modify the corresponding parameter of ArgumentParser(), then

python test_single.py

Quantitative experiments

  • PSNR.py provides three quantitative metrics - MSE, SSIM and PSNR - to allow you to observe the performance of inpainting.

python PSNR.py

Citation

@article{CHEN2021,
title = {TG-Net: Reconstruct visual wood texture with semantic attention},
journal = {Computers & Graphics},
year = {2021},
issn = {0097-8493},
doi = {https://doi.org/10.1016/j.cag.2021.09.006},
url = {https://www.sciencedirect.com/science/article/pii/S0097849321001928},
author = {Jiahao Chen and Yilin Ge and Quan Wang and Yizhuo Zhang},
keywords = {Wood defect treatment, Image inpainting, Generative adversarial networks, Attention mechanism}
}