Skip to content

Latest commit

 

History

History
67 lines (45 loc) · 6.06 KB

README.md

File metadata and controls

67 lines (45 loc) · 6.06 KB
Raw

ManTra-Net reproducing with pytorch

Powered by GitHub repo size GitHub Ask Me Anything ! visitors

Un-official inplementation of ManTra-Net which was proposed in CVPR 2019 by Yue Wu et al, includes code for training!.

What need to point out is that, the model in this Repo is mainly written by Pytorch, while the original paper provide a Keras-Tensorflow sample on Github.

Introduction

Files in the repo

  • Mantra_Net.py is the class of ManTra-Net which inherit from torch.nn.Module. This script imports several layers from directory ./imports, including:

    • CombindConv2D.py: The combination of 3 kinds of different layers:

      • BayarConv2D, first proposed by Bayar in 2018. Its filters can be illustrate as the figure below. Has a stronger ability to detecte manipulations.

      • SRMConv2D, first proposed in this paper and implemented in this paper. Three fixed filters are used to extract noise features. And when doing the process on RGB pictures, there will be 9 output channel for a 3 channel inputs.

      • Other filters are the normal Conv2D filters. In this reproduction, I choose (32 - 3 - 9) of this kind of filters, which is more than the number shows in the paper, but same as the orignal code on Github. (Really puzzling.)

    • ZPool2D.py An algorithm proposed in this ManTra-Net article. Aiming to quantify the difference between a local feature and a reference dominant feature.
    • convlstm.py An implementation of ConvLSTM layer, which was cited from ndrplz's Github repo.

  • What's more there is a ParametersManager.py file in ./imports dir. This is a part of code helps me to save parameters after tarinning for several batches, and load parameters then.

  • Train.py is the main script for model trainning on NIST16 dataset. If you want to run this file, you need to load NIST16 data in advance. Please read this NIST16 Dataset README for tutorials.

  • TestModel.py can illustrate the performance of the model (*.pt) you trained. It shows 4 sub-figures in a plt, including(From left to right): Raw picture, label mask, predict feature, predict mask.

  • TestModelAUC.py can calculate the AUC and ROC for you on trained models.

Some Comments

It should be noted that this model gives poor results if predicted without using the official pre-training parameters. The ROC curve below demonstrated that simply training and validating from the NIST16 dataset does not give good results.

However, we need to point out that in order to accelerate the trainning process, we resized images to (256x256), which reduced a lot of features, especially noise features. This problem has been discussed in the ManTra-Net paper as well. (Resize and JPEG compress will vitally do harm to the performance of the model)

We speculate that it is the strong dataset with 385 types of fine-grained manipulation images that made the model performs well. Sadly this dataset is not open to public :(

Links

If you want to train this Model with the CASIAv2 dataset, we provide a revised version of CASIAv2 datasets, which corrected several mistakes in the original datasets provided by the author. Details can find in the link shown below:

Readme Card

Cite

[1] Wu, Y., AbdAlmageed, W., & Natarajan, P. (2019). Mantra-net: Manipulation tracing network for detection and localization of image forgeries with anomalous features. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (pp. 9543-9552).

@InProceedings{Wu_2019_CVPR,
  author = {Wu, Yue and AbdAlmageed, Wael and Natarajan, Premkumar},
  title = {ManTra-Net: Manipulation Tracing Network for Detection and Localization of Image Forgeries With Anomalous Features},
  booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
  month = {June},
  year = {2019}
}

[2] Bayar, B., & Stamm, M. C. (2018). Constrained convolutional neural networks: A new approach towards general purpose image manipulation detection. IEEE Transactions on Information Forensics and Security, 13(11), 2691-2706.

[3] Fridrich, J., & Kodovsky, J. (2012). Rich models for steganalysis of digital images. IEEE Transactions on information Forensics and Security, 7(3), 868-882.