Processing-History-of-Images

Our aim is to reproduce the results mentioned in the paper Deep Learning for detecting Processing History of Images. We have used Pytorch for this implementation and have tried to stick to the CNN architecture used in the paper but have made certain tweaks to get better results.

Architecture

The training proceeds in two phases, and thus two CNN's are used, whose architectures are as follows:

Model I

Model II

Requirements

• python3.x
• latest version of Pytorch (you can follow any standard blog to install pytorch)
• numpy
• pickle
• virtualenv
• GPU with 10Gb or more memory(we used Nvidia 1080Ti for training)

Instructions

• Generating dataset:

  • First of all, BossBase raw images need to be downloaded which are present here.
  • Then, extract the tar files, create a folder datasets/1.data in the parent directory and copy the extracted images to this folder.
  • Next, open Matlab and run the commands cr2jpeg and then loldata within the matlab folder in the project. These commands might take quite some time.
  • After running these commands, the directory structure inside dataset folder will be as shown:

  .
  ├── jpegs
      ├── test
      |    |── ctr
      |    │    ├── denoise
      |    │    ├── high
      |    │    ├── low
      |    │    ├── org
      |    │    ├── tonal
      |    │
      |    |── mtr
      |         ├── denoise
      |         ├── high
      |         ├── low
      |         ├── org
      |         ├── tonal
      |       
      ├── train
      |    |── ctr
      |    │    ├── denoise
      |    │    ├── high
      |    │    ├── low
      |    │    ├── org
      |    │    ├── tonal
      |    │
      |    |── mtr
      |         ├── denoise
      |         ├── high
      |         ├── low
      |         ├── org
      |         ├── tonal        
      |        
      |
      |── val
           |── ctr
           │    ├── denoise
           │    ├── high
           │    ├── low
           │    ├── org
           │    ├── tonal
           │
           |── mtr
                ├── denoise
                ├── high
                ├── low
                ├── org
                ├── tonal        
  

• Training the network for phase 1:

  • Setting up a vitual environment:

  $ virtualenv venv
  $ source venv/bin/activate

  • Installing dependencies - pip install -r requirements.txt.
  • Training the model for phase I - python3.5 train_phase_1.py.
  • Run python3.5 test_phase_1.py to test the network.

• Once the network 1 is done training, we use the network to extract moments for images of random sizes. Run the following commands sequentially.

$ python3.5 extract_train_moments.py
$ python3.5 extract_val_moments.py
$ python3.5 extract_test_moments.py

• Training second part of the network:

  • Run python3.5 train_MLP_net.py to train second phase.
  • Run python3.5 test_MLP.py to test the network of second phase.

• Final detector is a cascade of both the networks.

The accuracies and other parameters used while training the models are mentioned in detail in the Report.

Contributors

• Aniket Singh
• Vishal Sharma