DiffusionModel

Environment configuration

• Python: 3.9
• Python library
  • OpenCV: 4.7.0.72
  • numpy: 1.23.5
  • scikit-image: 0.20.0
  • pytorch: 1.12.1 py3.9_cuda11.6_cudnn8_0
  • pyyaml: 6.0
  • torchvision: 0.5.0
  • dominate: 2.4.0
  • visdom: 0.1.8.8

Dataset structure

Model training requires paired stamp document images and real images, with a dataset structure of:

 dataset/
 ├── train/
 │   ├── input/
 │   │   ├── 1.png
 │   │   ├── 2.png
 │   │   └──...
 │   └── target/
 │       ├── 1.png
 │       ├── 2.png
 │       └──...
 ├── test/
 │   ├── input/
 │   │   ├── 1.png
 │   │   ├── 2.png
 │   │   └──...
 │   └── target/
 │       ├── 1.png
 │       ├── 2.png
 │       └──...

Run

Training

The model training parameters are configured in the configs.yml file. The key parameters are as follows:

• Training set path: train_data_dir
• Test set path: test_data_dir
• Output path: test_save_dir
• Intermediate result output path: val_save_dir
• Number of training epochs: n_epochs
• Weight path: resume
• Learning rate: lr
• Batch size: batch_size

train

python train_diffusion.py

test

python eval_diffusion.py

Data

The link to the dataset is dataset, and the link to the pre-trained model is weight.

In particular, since the complete document image contains sensitive information of many companies, we only provide the stamp area and the corresponding ground truth. We are trying to find a way to solve this problem. We are very sorry for the trouble caused to you. You can crop the stamp position of the initial document, process it and cover it to the position of the original image. In addition, since the dataset is randomly divided, the test results may not be exactly the same as in the paper, but generally will not exceed 0.1.