All Projects → Jongchan → Tensorflow Vdsr

Jongchan / Tensorflow Vdsr

A tensorflow implementation of "Accurate Image Super-Resolution Using Very Deep Convolutional Networks", CVPR 16'

Programming Languages

139335 projects - #7 most used programming language



This is a Tensorflow implementation for "Accurate Image Super-Resolution Using Very Deep Convolutional Networks", CVPR 16'.

  • The author's project page
  • To download the required data for training/testing, please refer to the at data directory.


  • : main training file.
  • : model definition.
  • : model definition for Factorized CNN. (not recommended to use. for record purpose only)
  • : define how to calculate PSNR in python
  • : test all the saved checkpoints
  • : plot the test result from

How To Use


# if start from scratch
# if start with a checkpoint
python --model_path ./checkpoints/CHECKPOINT_NAME.ckpt


# this will test all the checkpoint in ./checkpoint directory.
# and save the results in ./psnr directory

Plot Result

# plot the psnr result stored in ./psnr directory


The checkpoint is file is here

Results on Set 5
Scale Bicubic VDSR tf_VDSR
2x - PSNR/SSIM 33.66/0.9929 37.53/0.9587 37.24
3x - PSNR/SSIM 30.39/0.8682 33.66/0.9213 33.37
4x - PSNR/SSIM 28.42/0.8104 31.35/0.8838 31.09
Results on Set 14
Scale Bicubic VDSR tf_VDSR
2x - PSNR/SSIM 30.24/0.8688 33.03/0.9124 32.80
3x - PSNR/SSIM 27.55/0.7742 29.77/0.8314 29.67
4x - PSNR/SSIM 26.00/0.7027 28.01/0.7674 27.87


  • The training is further accelerated with asynchronous data fetch.
  • Tried to accelerate the network with the idea from Factorized CNN. It is possible to implement with tf.nn.depthwise_conv2d and 1x1 convolution, but not so effective.
  • Thanks to @harungunaydin 's comment, AdamOptimizer gives a much more stable training. There's an option added.
Note that the project description data, including the texts, logos, images, and/or trademarks, for each open source project belongs to its rightful owner. If you wish to add or remove any projects, please contact us at [email protected].