This is the repository of IMPRINTS, the idelible watermarking framework.

1 Preparation

1.1 Training and testing data

First, Download the colored large-scale watermark dataset (CLWD) to datasets/ from the link below:

Download the CLWD dataset: https://drive.google.com/file/d/17y1gkUhIV6rZJg1gMG-gzVMnH27fm4Ij/view?usp=sharing

Second, uncompress the CLWD.rar file datasets/, making datasets/CLWD/

1.2 Watermark removal models

Due to space limitation, we cannot upload the pretrain model for split-and-refine. You can download the pretrain model from the link below and place the 27kpng_model_best.pth.tar file in watermark_removal_works/split_then_refine.

The official repository of split-and-refine: https://github.com/vinthony/deep-blind-watermark-removal

2 Watermark optimization (training)

IMPEINTS supports two types of watermarks, logo and strings. For logo, it means you have to provide a png image that contains an alpha channel denoting the mask of the watermark. For string, it means you just have to provide a digital string input, IMPRINTS will convert the input into an image for you.

2.1 logo training

The main code for logo training is train_logo.py. Below is an example for training a logo in the CLWD-test, by specifying the path to the logo with --logo_path.

    python train_logo.py \
    --model slbr \
    --model_path ./watermark_removal_works/SLBR/pretrained_model/model_best.pth.tar \
    --output_dir ./adv_wm_pics/ \
    --logo_path ./datasets/CLWD/watermark_logo/test_color/1.png \
    --epoch 50 \
    --standard_transform 1 \
    --eps 0.2 \
    --log ./ckpt_wm/slbr/log/ \
    --gpu 0,1,2,3 \
    --batch_size 20

2.2 string training

The main code for string training is train.py. Below is an example for training a character or string input. The only difference is that the --logo_path is replaced by --text and --text_size.

    python train.py \
    --model slbr \
    --model_path ./watermark_removal_works/SLBR/pretrained_model/model_best.pth.tar \
    --output_dir ./adv_wm_pics/ \
    --text "IMPRINTS" \
    --text_size 50 \
    --epoch 50 \
    --standard_transform 1 \
    --eps 0.2 \
    --log ./ckpt_wm/slbr/log/ \
    --gpu 0,1,2,3 \
    --batch_size 20

3 Watermark deployment

Suppose you have trained a directory of logos or strings. Two formats are supported:

1. a directory of .png images, each one is a watermark, naive or IMPRITNS'
2. a directory of directories generated by IMPRINTS, i.e.,

└── 2022-12-XX-HH-MM-SS
    ├── wm_adv_0.pt
    ├── wm_adv_best.png
    ├── wm_adv_latest.png
    └── wm_adv_mask_end.pt

Then run deploy_many_logos.sh can deploy the watermarks on the host images of CLWD-test. P.S. specify the saving path first.

bash deploy_many_logos.sh \
    ${save_dir} \
    ./ckpt_wm/slbr/log/ \
    False \
    False \
    False \
    False \
    False \
    0.6 \
    1 \

See deploy_many_logos.sh and deploy_many_logos.py for detailed definitions of the parameters.

Char concatenation

You can use concat.py to concatenate you pre-trained characters into a .png image, such that you can use it as a logo for deployment afterwards.