Chroma wrinkles detector

Chroma wrinkles detector as described in "Image Forensics from Chroma Subsampling of High-Quality JPEG Images", IH&MMSec 2019.

Requirements

• DCT coefficient decoder for decoding DCT coefficients from JPEG-compressed images. Please follow the instructions from the GitHub repository.
• Python packages (available from PyPi):

numpy
pandas
tqdm
h5py
scipy
pyexiftool

• The exiftool command line tool must be installed.

Running the detector

This script takes a directory of JPEG-compressed images and computes the average correlation of each image's Cb and Cr channels with the chroma wrinkle template in DCT domain.

Note: Make sure that the decoder is included in your Python path.

Usage:

python classification/compute_scores_dct_matching.py
    [--quality QUALITY]
    [--reduce_444_chroma]
    [--crop]
    [--noise_residual]
    data_dir
    output_csv
    quality_factor_estimator_filename

Required arguments:

• data_dir: Directory in which to look for .jpg files (recursively).
• output_csv: Path to output csv file where to store results.
• quality_factor_estimator_filename: Path to HDF5 file that contains known quantization tables.

Optional arguments:

• quality: Restrict experiments to files matching to quality_{}.(jpg|jpeg)$.
• reduce_444_chroma: Boolean flag whether to subsample 4:4:4 images. Useful for images that were recompressed with no chroma subsampling.
• crop: Boolean flag whether to crop a random number of pixels from the top and left margins.
• noise_residual: Boolean flag whether to work on DCT coefficients of noise residual rather than decoded DCT coefficients.

Example:

cd classification
PYTHONPATH=~/i1/chroma-wrinkles:~/i1/dct-coefficient-decoder python compute_scores_dct_matching.py \
    /path/to/images \
    /tmp/output.csv \
    ../data/quality_factor_estimator_libjpeg_state.h5

Creating images with simple and DCT subsampling

Simple vs. DCT subsampling

Chroma wrinkles are introduced by libjpeg v6b and older, as well as noticeable forks such as libjpeg-turbo and mozjpeg. From libjpeg v7 on, DCT scaling became the default scaling operation. Therefore, newer versions of libjpeg do not introduce chroma wrinkles. Nevertheless, current versions of libjpeg resort to simple scaling instead of DCT scaling with the -nosmooth switch.

1. Update the paths to your local cjpeg and djpeg executables in utils/constants.py.

2. Store your RAW images in some directory.

• Download files.txt for raw images dataset, e.g., RAISE-1k.
• Download individual RAW files using wget -i files.txt.

3. Run create_data.py.

python create_data.py --encoders ENCODERS [ENCODERS ...]
                      [--quality_factors QUALITY_FACTORS [QUALITY_FACTORS ...]]
                      [--qtables QTABLES] [--sample SAMPLE]
                      input_dir output_dir

Required arguments:

• input_dir: Given this input directory, process all RAW images with the file extensions .nef or .dng.
• output_dir: Where to store resulting JPEG files.
• --encoders: List of encoders (1 or more) to feed data to.
• --quality_factors: List of quality factors (1 or more) for saving resulting images.

Optional args:

• --qtables: Encode with quantization table from given file path.
• --sample: HxV chroma subsampling

Example:

cd data
PYTHONPATH=~/i1/chroma-wrinkles python create_data.py \
    ~/data/RAISE_1k/raw/ \
    ~/data/RAISE_1k/ \
    --encoders libjpeg_dct_scaling libjpeg_turbo \
    --quality_factors 100 90 80

This will create one directory for each encoder in the given output folder.

The encoders pipe the output of dcraw into their respective implementation of cjpeg.

dcraw -w -c Ricoh_GX100_2_39129.DNG | cjpeg -quality 100 -outfile Ricoh_GX100_2_39129.jpg

The switch -w instructs dcraw to use camera whitebalance, and -c writes the output to stdout.