An open source code for paper "Distinguishing Computer-Generated Images from Natural Images Using Channel and Pixel Correlation" accepted in JCST 2020.

Example results of our method on images from the highly challenging dataset of Google versus PRCG. The demo lists the confidence of two labels (NI and CG) and calculates the time cost.

• Ubuntu 16.04
• Python 3
• NVIDIA GPU CUDA 9.0 + cuDNN 7.1.4
• PyTorch (tested on 0.4.1 and 1.0.0)
• Pillow (Python Imaging Library)
• Tqdm (Progress Bar for Python)

• Download the training and testing set: Columbia PRCG and SPL2018.
• Training: Set parameters in train.py or at the command line.

python train.py --batch-size 64 --patch-size 96 --epoch 1200 --lr 0.001

• Testing: Run python test.py.

@article{zhang2020distinguishing,
  title={Distinguishing Computer-Generated Images from Natural Images Using Channel and Pixel Correlation},
  author={Zhang, Rui-Song and Quan, Wei-Ze and Fan, Lu-Bin and Hu, Li-Ming and Yan, Dong-Ming},
  journal={Journal of Computer Science and Technology},
  volume={35},
  pages={592--602},
  year={2020},
  publisher={Springer}
}

This code refers to pytorch-CycleGAN-and-pix2pix and a Caffe version NIvsCG.