SRDM — Super-Resolution Detection Method

SRDM is a new upscaled-resolution-detection method based on learning of visual representations using contrastive and cross-entropy losses.

See our paper for details.

Dependencies and Installation

• Python >= 3.8 (Recommend to use Anaconda or Miniconda)
• PyTorch >= 1.8

Installation

1. Clone repo

   git clone https://github.com/msu-video-group/SRDM.git
   cd SRDM

2. Install dependent packages

    pip install -r requirements.txt

Quick Inference

1. Download pre-trained models from Google Drive. Place the models in ./pretrained_models. We provide two models: MobileNet and ResNet.
2. Place your own videos as:

  $ tree data
  ├── data
      ├── video_001
      │   ├── frame_001.png
      │   ├── frame_002.png
      │   └── ...
      ├── video_002
      │   ├── frame_001.png
      │   ├── frame_002.png
      │   └── ...
      └── ...

3. We provide a criterion for fake-resolution video. The method returns the probability of fake resolution for each frame. Then if 0.1 quantile of probabilities is bigger than 0.5 the video will be considered as fake-resolution video. You can config your own criteron in the test.py
4. Run test.

   python3 test.py --pretrained=<checkpoint> --test-data-path=<data path> --n_frames=<> --embedding-size=<> --model=<>

5. For example.

   python3 test.py --pretrained=./pretrained_models/resnet-emb-64-n_fr-3.pt --test-data-path=<data path> --n_frames=3 \
           --embedding-size=64 --model=resnet

Description

SRDM is initially described in paper. It's a novel approach to video-super-resolution detection that combines contrasting and supervised learning. The method first uses a ResNet backbone to extract deep features and then uses a lightweight MLP networks to project and classify representation of input frames. Our method achieved good performance in extensive experiments.

Main Results on MSU Video Super-Resolution Benchmark

SR method	Accuracy		SR method	Accuracy
Real-ESRGAN	0.875		Real-ESRnet	1.000
DynaVSR-R	1.000		DynaVSR-V	1.000
D3Dnet	1.000		DBVSR	1.000
DUF-16L	1.000		DUF-28L	1.000
ESPCN	0.875		LGFN	1.000
RRN-10L	1.000		RRN-5L	1.000
SOF-VSR-BD	1.000		SOF-VSR-BI	1.000
TDAN	1.000		RealSR	0.875
ESRGAN	0.625		TGA	1.000
RSDN	1.000		RBPN	1.000
iSeeBetter	1.000		TMNet	1.000
GFPGAN	1.000		SwinIR	0.875
SRMD	1.000		comisr	1.000
waifu2x-cunet	1.000		waifu2x-anime	1.000
BasicSR	1.000		VRT	1.000

Training

Data preparation

  $ tree data
  ├── train
  │   ├── original
  │   │   ├── GT
  │   │   │   ├── video_001
  │   │   │   │   ├── frame_001.png
  │   │   │   │   └── ...
  │   │   │   ├── video_002
  │   │   │   │   ├── frame_001.png
  │   │   │   │   └── ...
  │   │   │   └── ...
  │   │   ├── SR method 1
  │   │   │   ├── video_001
  │   │   │   │   ├── frame_001.png
  │   │   │   │   └── ...
  │   │   │   ├── video_002
  │   │   │   │   ├── frame_001.png
  │   │   │   │   └── ...
  │   │   │   └── ...
  │   │   └── ...
  │   └── compressed
  │   │   │   ├── video_001
  │   │   │   │   ├── frame_001.png
  │   │   │   │   └── ...
  │   │   │   ├── video_002
  │   │   │   │   ├── frame_001.png
  │   │   │   │   └── ...
  │   │   │   └── ...
  │   │   ├── SR method 1
  │   │   │   ├── video_001
  │   │   │   │   ├── frame_001.png
  │   │   │   │   └── ...
  │   │   │   ├── video_002
  │   │   │   │   ├── frame_001.png
  │   │   │   │   └── ...
  │   │   │   └── ...
  │   │   └── ...
  └── val
      └── ... 

Training from scratch

To train a SRDM on your data from scratch, change path to your dataset and run:

python3 main_worker.py --num_workers=2 --n_frames=2 --model-name=resnet --batch-size=32 

Cite Us

@article{meshchaninov2022combining,
  title={Combining Contrastive and Supervised Learning for Video Super-Resolution Detection},
  author={Meshchaninov, Viacheslav and Molodetskikh, Ivan and Vatolin, Dmitriy},
  journal={arXiv preprint arXiv:2205.10406},
  year={2022}
}