Awesome

Square Images Colorization

This repo contains a project focused on the developing of a colorization system for images depicting (not only) cities squares. The idea is to produce two different models: the first based on an autoencoder structure and the second gans-based.

Setup the deep learning environment

Setup the virtual environment for the project

$ conda create -n venv python=3.7.4 anaconda

Add the virtualenv as a jupyter kernel

$ ipython kernel install --name "image_colorization" --user

Install Pytorch.

$ conda install pytorch torchvision cudatoolkit=10.0 -c pytorch

Change cudatoolkit preference if you have another CUDA version. If you don't have gpu:

$ conda install pytorch torchvision cpuonly -c pytorch

Install requirements

$ pip install -r requirements.txt

PROJECT ROADMAP

The first part of this project consists of "data acquisition" regarding images of cities squares. To this intend a script to scrape images from google image has been developed. To use it, place get_data.py and url.yaml files in the same directory and specify the google image page url you intend to scrape in url.yaml.

$ python get_data.py --dest_folder [name of destination folder]

Furthermore, Instagram-Scraper has been used to get more data and search them by the hashtag. Hashtag searched are piazzedimilano, piazzeditalia and citysquares.

- AutoEncoder Results

- Pix2pix

The gan's based approch relies on Pix2pix architecture, following this implementation. Other images have been captured with Instagram scraper and added to the training dataset in order to improve the model generalization capability. The model has been trained for 105 epochs with a learning rate of 0.0002 and Adam optimizer. For other training details, please refer to the original implementation. Below, it's possible to look at some results:

The Pix2pix model has also been tested on the same images tested on Autoencoder previously developed.

Looking at these images we can say the the Autoencoder performs better on vintage photo while Pix2pix gets betters results, not only on cities squares images, but also on landscapes and so on, giving lively colors to the input.