Double Compression Detection for Same Quantization Matrix using Multi-Coloumn CNN.

This repository contains the code for reproducing results presented in the paper "Double JPEG Compression Detection of Distinguishable Blocks in Images Compressed With Same Quantization Matrix" (MLSP 2020)

Environments and Dependenices

• tensorflow = '1.12.0' (cudatoolkit=9.0 & cudnn=7.1.2)
• scipy = '1.4.1'
• sklearn = '0.22.1'
• matplotlib = '3.1.3'
• numpy = '1.18.1'

We recommend installing the dependencies using conda.

Getting Started

Data Generation

• Download the UCID dataset containing 1338 images .TIF images and place it in the ./data folder in the root directory. Also download the jpeg-read-toolbox and place it in /code/data_creation/dependencies/ as jpeg_read_toolbox.

• Create another folder called all_needed_matlab_functions in the folder ./code/data_creation/.

• Then execute ./code/data_creation/data_maker.m to create the Compressed_UCID_gray_full dataset.

• Execute ./code/data_creation/patch_maker.m , to create the training and testing 8 x 8 patches, with both train = True and train = False.

• After this please execute ./code/data_creation/save_error_images.m in order to create all the error images for the generated patches.

The file structure after saving the dataset would be as follows,

/data
  /dataset
     /8
       /train
        /Quality_{Qf}
          /index_all
            /single
               /1
                 /single_error.mat
                 /single_dct_error.mat
               /2
                 ....
            /double
              /1
                 /double_error.mat
                 /double_dct_error.mat
              /2
                 ....
     

The dataset of error images is also available directly for download here, which is to be saved in ./data folder in the root directory.

Training and Inference

The main code to run our approach is MCCNN_Classifier.py, which use the default runtime settings of our model as mentioned in the paper. They include,

Command Line Parameters:

These parameters can be supplied at the time of running the code. Below are the possible combinations that can be given for command line arguments.

• Quality factor --Qf = {20,40,60,70,75,80,85,90}
• Stability index --index = {1 or all}
• Number of error images --stack ={2,3}
• Run for all quality factors --all_Q = {0,1}
• Number of repeated runs, --runs = {1,2,3,4,5,6,7,8,9,10}

Pretrained Models.

• For direct inference the pretrained models can be downloaded from here.
• To use each quality factor separately download and place it under proposed_results/MCNN/index_all.
• Alternatively, you can download all pretrained models by unzipping the entire folder in proposed_results/MCNN.

Sample Command.

To reproduce our results at --Qf = 60, --index = 'all' and --stack = 3 for --runs=10 the code is:

python /code/generate_results/MCCNN.py \
-- Qf = 60 \
-- index = all\
-- stack = 3\
-- runs = 10\
-- ptr = 1

• To execute for all quality factors please set the parameter --all_Q=1
• Set --ptr = 1 to use the pretrained model.

Obtaining Results

The results are saved in the following directory structure, with results_itr_~.mat containing the resultant predictions.

/proposed_results
  /MCNN
     /index_all
         /Quality_~
            /runs
              /MCNN_stack_3_scale_1
                  /results/
                      /results_itr_0.mat
                      ..
                      ..
                      ..
                  /loss_plots/
                      /loss_itr_0.png
                      ..
                      ..
                      ..
                  /weights/
                     /best_weights_itr_0.mat
                     ..
                     ..
                     ..

References:

• Schaefer, Gerald, and Michal Stich. "UCID: An uncompressed color image database." Storage and Retrieval Methods and Applications for Multimedia 2004. Vol. 5307. International Society for Optics and Photonics, 2003.

Citation

  title={Double JPEG Compression Detection for Distinguishable Blocks in Images Compressed with Same Quantization Matrix},
  author={Abhinav Narayan Harish*, Vinay Verma*, Nitin Khanna},
  booktitle={IEEE International Workshop on Machine Learning for Signal Processing}
  year={2020}