pyF2O

Forged Image To Original Image Generation

Version: 3.0.0   
Author : Md. Nazmuddoha Ansary
         Shakir Hossain  
         Mohammad Bin Monjil  
         Habibur Rahman
         MD.Aminul Islam
         Shahriar Prince  

Version and Requirements

• numpy==1.17.4
• tensorflow==2.0.0
• Python == 3.6.8

Create a Virtualenv and pip3 install -r requirements.txt

DataSet

1. Download Data Sets: MICC-F2000 and MICC-F220 dataset
2. Unzip MICC-F2000.zip FOR TRAINING and MICC-F220 FOR TESTING
   The MICC-F2000 dataset contains a file named: nikon7_scale.jpg. It has to be renamed as nikon_7_scale.jpg.

Preprocessing

config.json Change The following Values in config.json

    "ARGS":
    {
        "MICC-F2000"        : "/home/ansary/RESEARCH/F2O/UNZIPPED/MICC-F2000/",
        "MICC-F220"         : "/home/ansary/RESEARCH/F2O/UNZIPPED/MICC-F220/",
        "OUTPUT_DIR"        : "/home/ansary/RESEARCH/F2O/"
    }

clear_mem.sh (Ubuntu/Linux) The complete preprocessing may take huge time and also cause to crash the system due to high memory useage. A way around is built for Ubuntu users is to run sudo ./clear_mem.sh in parallel with main.py

        usage: main.py [-h] exec_flag

        Preprocessing Script:Forged Image To Original Image Reconstruction

        positional arguments:
        exec_flag   
                                                Execution Flag for creating files 
                                                Available Flags: png,tfrecords,comb
                                                png       = create images
                                                tfrecords = create tfrecords
                                                comb      = combined execution
                                                PLEASE NOTE:
                                                For Separate Run the following order must be maintained:
                                                1) png
                                                2) tfrecords
                                                
                                                

        optional arguments:
        -h, --help  show this help message and exit

Results

• If execution is successful a folder called DataSet should be created with the following folder tree:

        DataSet  
        ├── test
        │   ├── image
        │   └── target
        ├── tfrecord
        │   ├── test
        │   └── train
        └── train
            ├── image
            └── target
  

ENVIRONMENT

OS          : Ubuntu 18.04.3 LTS (64-bit) Bionic Beaver        
Memory      : 7.7 GiB  
Processor   : Intel® Core™ i5-8250U CPU @ 1.60GHz × 8    
Graphics    : Intel® UHD Graphics 620 (Kabylake GT2)  
Gnome       : 3.28.2  

GCS

Training with tfrecord is not implemented for local implementation. For using colab, a bucket must be created in GCS and connected for:

• tfrecords
• checkpoints

Networks

• Generator structre

• Discriminator structre

pix2pix

Image Source

Original paper: Image-to-Image Translation with Conditional Adversarial Nets

Implementation based on official tensorflow tutorial

• run pix2pix_gpu.ipynb in colab