Double JPEG Detection in Mixed JPEG Quality Factors using Deep Convolutional Neural Network (ECCV 2018)

PyTorch version (2021-02-20)

Please move to below repository.

Title: Double JPEG Detection in Mixed JPEG Quality Factors using Deep Convolutional Neural Network
Conference: The European Conference on Computer Vision (ECCV)
Data normalization is important. You should set the Y value having 0-255, not 0-1.
Proposed network tensorflow code.
Trained weights link: https://drive.google.com/open?id=1HlKSoCGVm_2Gia--MH5Bx3SquJos-kr1.

The network was trained with two dataset.

First dataset was generated using 1120 quantization tables (mentioned in the paper).

Second dataset was generated using standard quantization tables (Q50 - Q100).
Example code (application.py) is updated now.

normal_galaxy_note8_1.jpg - normal image was captured by galaxy note8 smartphone.

splicing_PQ10.jpg - manpulated image (splicing) was generated by Photoshop.

copy_move_PQ10.jpg - manpulated image (copy-move) was generated by Photoshop.

PQ# means save jpg quality factors in Photoshop.
You can download single and double JPEG images and 1120 quantization tables at https://sites.google.com/view/jspark/home/djpeg.
Videos showing the process of generating manipulated images. https://drive.google.com/file/d/1G3fOUIBgI3Th0bjPVvP6Dzba-o_Wx3Ev/view?usp=sharing
Rebuttal. https://drive.google.com/file/d/1RFLk0b_pilQUYpz82mTFOl0aR-Q-IfSA/view?usp=sharing