Awesome

Example-based Colorization via Dense Encoding Pyramids

1552888570716

Example-based Colorization via Dense Encoding Pyramids, Chufeng Xiao, Chu Han, Zhuming Zhang, Jing Qin, Tien-Tsin Wong, Guoqiang Han, Shengfeng He, Computer Graphics Forum, 2019.

Prerequisites

Linux
Caffe & Pycaffe
Python 2.7
Python libraries (numpy, skimage, scipy)

Getting Started

Compile Caffe

Note: If you don't need to train this model (only test it), you can skip step 1 and 2 below and directly compile the original version of caffe, i.e., you don't have to copy and paste the files mentioned in step 1 and 2.

copy two files softmax_cross_entropy_loss_layer.cpp and softmax_cross_entropy_loss_layer.cu under the folder ./resources into <your caffe path>/caffe/src/caffe/layers/
copy the file softmax_cross_entropy_loss_layer.hpp under the folder ./resources into <your caffe path>/caffe/include/caffe/layers

Note that you also need to compile pycaffe and add it into your PYTHONPATH:

vi ~/.bashrc

# add the two lines into the file
PYTHONPATH=<you caffe path>/caffe/python:$PYTHONPATH
LD_LIBRARY_PATH=<you caffe path>/caffe/build/lib:$LD_LIBRARY_PATH

# save and update the environment
source ~/.bashrc

compile and test caffe:

# execute under the root directory of caffe
make clean # clean the files complied before

make all
make test

Add Interface Files into Settings

In order to use the interface files for caffe layer, you need to add the path of the folder ./resources

vi ~/.bashrc

# add this line
export PYTHONPATH=$PYTHONPATH:~/<DEPN path>/resources

# save and update the environment
source ~/.bashrc

Download the Models of DEPN

There are two models you need to download for testing or training. DEPN_init.caffemodel saves the first-level parameters of DEPN, while DEPN_sub.caffemodel provides the shared parameters used by the second level and over. For convenient connection, we provide both of OneDrive and Google Drive links for downloading. Please put the two models under the file folder ./models.

Google Drive links
- DEPN_init.caffemodel
- DEPN_sub.caffemodel
Alternatively, OneDrive links:
- DEPN_init.caffemodel
- DEPN_sub.caffemodel

Test and Generate Colorful Images

You can choose any image as a reference for the grayscale image, even a palette. Just simply execute test.py :

python test.py -gray <gray_dir> -refer <refer_dir> -output <output_dir>

# Example
python test.py -gray ./test_img/gray/1.jpg -refer ./test_img/refer/1.jpg -output ./test_img/result/1.png

Please make sure the size of the grayscale image is at least 64*64. If you want to test the image with smaller size or want to adjust the first-level input size of DEPN, you should change the value of init_levelin test.py to the size you desire. And then create a new file DEPN_deploy_<size>.prototxt:

copy and paste the file DEPN_deploy_64.prototxt under the ./models/test/
change the name of the new file to DEPN_deploy_<new_size>.prototxt

edit the fileDEPN_deploy_<new_size>.prototxt and change all the values 64 of the input layer to the new size:

layer {
  name: "img_l"
  type: "Input"
  top: "img_l"
  input_param {
    shape { dim: 1 dim: 1 dim: 64 dim: 64 }
  }
}

layer {
  name: "ref_ab"
  type: "Input"
  top: "ref_ab"
  input_param {
    shape { dim: 1 dim: 2 dim: 64 dim: 64 }
  }
}

The procedures of changing the input size of the second level and over are similar to these.

Training

Prepare Dataset

You need to transform the dataset of images into LMDB files, which can be used for training through caffe.

For the first level of DEPN, you should only prepare a LMDB file with a set of colorful images, which will be automatically divided into grayscale image, namely luminance channel, and ground truth by our codes.

For the levels above the first, all of them not only require images with the corresponding size as datasets, like the first level, but also need the small outcomes from the former level, which means that you should generate two LMDB files. You can use the codes in test.py to get the small outcome at the former level.

....

# if you need to use the small outcome to train the higher levels, please use the codes below:
small_img_rgb=caffe.io.resize_image(img_rgb,(size/4,size/4))
small_img_lab = color.rgb2lab(small_img_rgb)
small_img_l = small_img_rgb[:,:,0]
small_img_lab_out = np.concatenate((small_img_l[:,:,np.newaxis],ab_dec),axis=2)
small_img_rgb_out = (255*np.clip(color.lab2rgb(small_img_lab_out),0,1)).astype('uint8')
scipy.misc.toimage(small_img_rgb_out).save(sm_out)

....

Edit Network Prototxt

After getting the LMDB files, you should edit the network prototxt, like ./models/train/DEPN_64.prototxt and ./models/train/DEPN_128.prototxt, to place the path of your LMDB files as the value of source.

If you want to change the input size of DEPN while training, you also can change the sizes of the images in LMDBs and correspondingly replace the value of crop_size.

layer {
  name: "data"
  type: "Data"
  top: "data"
  include {    phase: TRAIN  }
  transform_param {
   mirror: true
   crop_size: 64
  }
  data_param {
    source: "" # [[REPLACE WITH YOUR PATH]]
    batch_size: 5
    backend: LMDB
  }
}

layer {
  name: "data"
  type: "Data"
  top: "data"
  include {    phase: TEST  }
  transform_param {
   mirror: true
   crop_size: 64
  }
  data_param {
    source: "" # [[REPLACE WITH YOUR PATH]]
    batch_size: 1
    backend: LMDB
  }
}

Edit Training Prototxt

Before starting training, please change the net position, i.e., the path of network prototxts, in the file ./models/train/solver.prototxt.

Start Training

Execute sh ./models/train/train_DEPN.sh to start training. You maybe need to change the caffe position to match that in your machine. And if you want to train the network based on our models, you can set ./models/DEPN_init.caffemodel or ./models/DEPN_sub.caffemodel as a pre-trained model.

<Your install path>/caffe/build/tools/caffe train -solver ./models/train/solver.prototxt -gpu 0 -weights ./models/DEPN_init.caffemodel

Acknowledgement

Part of the code is based on Colorful Image Colorization.