Find file History
Latest commit c706b2a Dec 18, 2015 rensijie Default data has only 1 channel. Fix the reshape issue for this defau…
…lt setting.
Permalink
..
Failed to load latest commit information.
Shepard_inpainting
Shepard_super_res
utility
README.md Update README.md Dec 7, 2015

README.md

Shepard Convolutional Neural Networks

Before running the code, it's a good idea to read the instructions and see the videos on the VCNN website (www.deeplearning.cc/vcnn) and setup VCNN.
You need to have a Nvidia GPU with 4GB or more GPU memory to run the code. The code runs directly in Matlab 2014b or later. You need to install CUDA before running the code. No other compilation or configuration is needed. The code was tested in both Ubuntu 14.04 and Windows 7.

Paper

Jimmy SJ. Ren, Li Xu, Qiong Yan, Wenxiu Sun, "Shepard Convolutional Neural Networks", Advances in Neural Information Processing Systems (NIPS 2015)

Super-resolution

We provided the pre-trained models for 2x, 3x and 4x super-resolution.
For PSNR comparison, following previous papers in the literature, we used single channel images.
Please run this script to test 2x super-resolution.
Please run this script to test 3x super-resolution.
Please run this script to test 4x super-resolution.
Most of the methods we compared against can be found in this web page.

We also provided scripts to generate color images.
Please run this script to test 2x color image super-resolution.
Please run this script to test 3x color image super-resolution.
Please run this script to test 4x color image super-resolution.

To train a new model, run this script to generate the data. You can adjust the variable "down_factor" to indicate whether you would like to generate data for x2, x3 or x4 super-resolution. Also adjust the variable "training_mat_path" and "val_mat_path" to indicate the path to store the training and validation data.

Once the data is ready, you may run this script to train a x2 super-resolution model. This script for x3 super-resolution model. This script for x4 super-resolution model.

See a few examples for color super-resolution.
x4 super-resolution

Bicubic VS. Shepard CNN

Bicubic VS. Shepard CNN

x3 super-resolution

Bicubic VS. Shepard CNN

x2 super-resolution

Bicubic VS. Shepard CNN

Inpainting

To run examples of inpainting just run this script.

See a few examples for image inpainting.