Skip to content
Zhicheng Yan edited this page May 17, 2015 · 44 revisions

Welcome to the dl-image-enhance wiki!

Introduction

dl-image-enhance project implements algorithms described in [1].

It should be used along with another project cuda_convnet_plus to reproduce the experimental results reported in [1].

Setup

  • Adapt ${DL_IMAGE_ENHANCE_ROOT}/setup.sh.example. Set environment variable PROJ_DIR properly
  • Rename it to setup.sh. Run the script ". setup.sh" to set up the environment.

Local effect dataset

There are three local effect datasets, namely Foreground Pop-out, Local Xpro and Watercolor.

Original images path : dl-image-enhance/data/uniform_set/uniform_set_autotone_tif

Foreground Pop-out images path: data/uniform_set_foregroundpopout/uniform_set_foregroundpopout_tif

Local Xpro images path: data/uniform_set_xpro/uniform_set_xpro_tif

Watercolor images path: data/uniform_set_watercolor/uniform_set_watercolor_tif

Demo: Foreground Pop-out

We prepare scripts to reproduce the results of Foreground Pop-out effect in the paper [1]. Please sequentially execute the following scripts.

  • Prepare training data batches for neural network training.

Script: ${DL_IMAGE_ENHANCE_ROOT}/py/script_prepare_train_batches_uniform_set_foregroundpopout.sh


  • Optionally, train a deep net. Otherwise, use the a provided trained model to enhance image at testing stage.

Script: ${CUDA_CONVNET_PLUS_ROOT}/py/script_train_uniform_set_foregroundpopout.sh


  • Use a pre-trained neural network to enhance testing images. The unit of enhancement is super-pixel. The enhanced testing images will be saved in training checkpoint summary folder ${DL_IMAGE_ENHANCE_ROOT}/data/uniform_set_foregroundpopout/convnet_checkpoints/ConvNet__2015-01-06_20.11.39zyan3-server2_summary/superpixel

Script: ${DL_IMAGE_ENHANCE_ROOT}/py/script_enhance_images_uniform_set_foregroundpopout.sh


Demo: Local Xpro

Please sequentially execute the following scripts to reproduce results of Local Xpro effect.

  • ${DL_IMAGE_ENHANCE_ROOT}/py/script_prepare_train_batches_uniform_set_xpro.sh
  • ${CUDA_CONVNET_PLUS_ROOT}/py/script_train_uniform_set_pro.sh
  • ${DL_IMAGE_ENHANCE_ROOT}/py/script_enhance_images_uniform_set_xpro.sh.

The enhanced testing images will be saved in training checkpoint summary folder ${DL_IMAGE_ENHANCE_ROOT}/data/uniform_set_xpro/convnet_checkpoints/ConvNet__2015-01-06_14.45.07zyan3-server2_summary/superpixel

Demo: Watercolor

Please sequentially execute the following scripts to reproduce results of Watercolor effect.

  • ${DL_IMAGE_ENHANCE_ROOT}/py/script_prepare_train_batches_uniform_set_watercolor.sh
  • ${CUDA_CONVNET_PLUS_ROOT}/py/script_train_uniform_set_watercolor.sh
  • ${DL_IMAGE_ENHANCE_ROOT}/py/script_enhance_images_uniform_set_watercolor.sh.

The enhanced testing images will be saved in training checkpoint summary folder ${DL_IMAGE_ENHANCE_ROOT}/data/uniform_set_watercolor/convnet_checkpoints/ConvNet__2015-01-06_14.45.52zyan3-server2_summary/superpixel

References

  1. Zhicheng Yan, Hao Zhang, Baoyuan Wang, Sylvain Paris, Yizhou Yu. Automatic Photo Adjustment Using Deep Neural Networks. ACM Transactions on Graphics (TOG), 2015