Automatic Image Colorization using CNNs and InceptionResnetV2

Overview

This project showcases my implementation of Baldassarre et al.'s Deep Koalarization: Image Colorization using CNNs and Inception-ResNet-v2 research paper from 2017 using PyTorch. The network is trained using 60 000 images from ImageNet. You can find more information in the Jupyter notebook.

Results

More results can be found in the notebook and the image results folder.

Steps to Run

1. Run pip install -r requirements.txt to install the necessary dependencies for both training and predicting.
2. To retrain the model using your own dataset, run the notebook and replace the file paths with your own.
3. To colourize your own images using the model provided in models, run streamlit run app.py. The following UI should be accessible at localhost:8501.

Note: The model's image size output is 224x224. If the grayscale image is larger than this, it will be centre cropped to fit these dimensions. To prevent the image from being cut off, resize the image to fit these dimensions before running the prediction script.

Built With

• PyTorch
• Pretrained-Models.PyTorch - Leveraged InceptionResnetV2
• ImageNet Downloader - Downloading the dataset

Credits

Here are the implementations that gave me inspiration for this project:

• deep-koalarization - baldassarreFe
• Automatic-Image-Colorization - lukemelas
• hands-on-transfer-learning-with-python - dipanjanS

Author

Laura Dang