image_colorization

image Colorization is the practice of estimating color channels for grayscale images or video frames. In this project, we make use of CIELAB color space to estimated the a*, b* components given the luminance component. For this project, we implemented three models, which are as follow:
1- Deep Koalarization: Image Colorization using CNNs and Inception-V3
2- Attention U-Net
3- Fusion of Attention U-Net model with Inception-V3 as a global feature extractor

Setup

$ python3 -m venv ~/.virtualenvs/env
$ source ~/.virtualenvs/env/bin/activate

Updating project dependencies

# And to install the packages
$ pip install -r requirements.txt

Getting the dataset

Places 205 dataset Data

Running the code

Default arguments

EPOCHS=100
BATCH_SIZE=32
LEARNING_RATE=1e-3

Training from scratch

1- Deep Koalarization model
python main.py --epochs=100 --lr=0.0005 --seed=5 --batch_size=32 --model_name=koalarization
2- Attention U-Net
python main.py --epochs=100 --lr=0.0005 --seed=5 --batch_size=32 --model_name=attention_unet
3- Fusion of Attention U-Net model with Inception-V3
python main.py --epochs=100 --lr=0.0005 --seed=5 --batch_size=32 --model_name=attention_unet_fusion

Results

Some of the colored samples from the test set are shown below. The images in the first column show original images, the images in the second column denote colored images by Attention U-Net model, and the images in the third column denote colored images by Deep koalarization model.

License

MIT License

shakibyzn/image_colorization