DeepMIH: Deep Invertible Network for Multiple Image Hiding (TPAMI 2022)

This repo is the official code for

Published on IEEE Transactions of Pattern Analysis and Machine Intelligence (TPAMI 2022).
@ Beihang University.

1. Pre-request

1.1 Dependencies and Installation

1.2 Dataset

  • In this paper, we use the commonly used dataset DIV2K, COCO, and ImageNet.
  • For train or test on your own path, change the code in
    line50: TRAIN_PATH_DIV2K = ''
    line51: VAL_PATH_DIV2K = ''
    line54: VAL_PATH_COCO = ''
    line55: TEST_PATH_COCO = ''
    line57: VAL_PATH_IMAGENET = ''

2. Test

  1. Here we provide a trained model.
  2. Download and update the MODEL_PATH and the file name suffix before testing by the trained model.
    For example, if the model name is,,,
    and its path is /home/usrname/DeepMIH/model/,
    PRETRAIN_PATH = '/home/usrname/DeepMIH/model/',
    PRETRAIN_PATH_3 = '/home/usrname/DeepMIH/model/',
    file name suffix = 'model_checkpoint_03000'.
  3. Check the dataset path is correct.
  4. Create an image path to save the generated images. Update TEST_PATH.
  5. Run

3. Train

  1. Create a path to save the trained models and update MODEL_PATH.
  2. Check the optim parameters in is correct. Make sure the sub-model(net1, net2, net3…) you want to train is correct.
  3. Run Following the Algorithm 1 to train the model.
  4. Note: DeepMIH may be hard to train. The model may suffer from explosion. Our solution is to stop the training process at a normal node and abate the learning rate. Then, continue to train the model.

4. Further explanation

In the at line 223:
rev_secret_dwt_2 = rev_dwt_2.narrow(1, 4 * c.channels_in, 4 * c.channels_in) # channels = 12,
the recovered secret image_2 is obtained by spliting the middle 12 channels of the varible rev_dwt_2. However, in the forward process_2, the input is obtained by concatenating (stego, imp, secret_2) together. This means that the original code has a bug on recovery process (the last 12 channels of the varible rev_dwt_2 should be splited to be the recovered secret image_2, instead of the middle 12 one). We found that in this way the network is still able to converge, thus we keep this setting in the test process.
We also offer a corrected version (see line 225) and You can also train your own model in this way.

Feel free to contact: [email protected].


If you find this repository helpful, you may cite:

  author={Guan, Zhenyu and Jing, Junpeng and Deng, Xin and Xu, Mai and Jiang, Lai and Zhang, Zhou and Li, Yipeng},
  journal={IEEE Transactions on Pattern Analysis and Machine Intelligence}, 
  title={DeepMIH: Deep Invertible Network for Multiple Image Hiding}, 


View Github