zengh5 / ISONet

IWDW2020: ISO Setting Estimation Based on Convolutional Neural Network

ISONet

ISONet: ISO Setting Estimation Based on Convolutional Neural Network and Its Application in Image Forensics

Abstract

The ISO setting, which is also known as film speed, influences the noise characteristics of output images. As a consequence, it plays an important role in noise based forensics. Whenever the ISO setting information cannot be retrieved from the image metadata, estimating the ISO setting of a probe image from its content is of forensic significance. In this work, we propose a convolutional neural network, called ISONet, for ISO setting estimation. The proposed ISONet can successfully infer the ISO setting both globally (image-level) and locally (patch-level). It not only work on uncompressed images, but also is effective on JPEG compressed images. We apply the ISONet on two typical forensic scenarios, one is the image splicing localization and the other is the Photo Response Non-Uniformity (PRNU) correlation prediction. A series of experiments show that the ISONet can yield a remarkable improvement in both forensic scenarios.

To appear in IWDW2020

Network structure

Dependencies

From running our demo,following packages are required

• pytorch >=1.3
• opencv-python
• matplotlib
• h5py
• numpy
• tensorboard

Test ISONet

Test ISONet_Uncompressed

Run python test_model.py --model_path models/net_tif.pth --pic_path image_path
or
Run Fig4.py to repeat the ISO metric map used for Fig.4

Test ISONet_JPEG

Runpython test_model.py --model_path models/net_jpg.pth --pic_path image_path
or
Run Fig7.py to repeat the global ISO metric used for Fig.7

Train ISONet

Take ISONet_Uncompressed for example

1. Download the training images to IMG_DATA_DIR of your computer

2. runprepare_data.py to generate the training set

   python prepare_data.py --aug_times 3 --pic_type tif --data_path IMG_DATA_DIR --save_path DATA_PATH

3. Run main_train.py to train the model python main_train.py --data_path DATA_PATH

About the author

Most of the codes are provided by Kang Deng, [e-mail]