/ImageSR-Tensorflow

An image super-resolution implemented by Tensorflow

Primary LanguagePythonMIT LicenseMIT

ImageSR for Tensorflow

MIT License

Update

Preparation

  • Python 2.7/3.4/3.5/3.6 (recommend python 3.5/3.6)
  • Tensorflow 1.4+ (utilize tf.data to read training data)

Training

  • Prepare training dataset. Please download DIV2K dataset: link
  • Move the training dataset to ./DIV2K/DIV2K_train_HR
  • Check the config in train.py
  • Run train.py > python train.py

Inference

We train two models: VDSR (only x4 scale) and EDSR (small version). Models are trained 50,000 steps with batch size 64.
Please download model ckpt files from Cloud:
BaiduYun:
link：https://pan.baidu.com/s/1eB0jY5cgp7j39M3Gt2ndqw password：4qm4
GoogleDrive:
line: https://drive.google.com/open?id=1IUKH4f0LzkobU4zktA3ZzEg86-ufrV1c

  • Check the config in test.py
  • Run test.py > python test.py

Performance

Using our models can achieve the performance as shown in below table data:

Dataset Scale Bicubic VDSR EDSR
Set5 x4-PSNR 28.43 31.43 31.90

x4 SR Visual Result

BICUBIC VDSR EDSR
BICUBIC VDSR EDSR