Skip to content

Latest commit

 

History

History
78 lines (47 loc) · 4.71 KB

README.md

File metadata and controls

78 lines (47 loc) · 4.71 KB
Raw

Thank you for visiting this repository

This repo is the artefact of research project "Experiments on Restoring Color and Visual Fidelity to Legacy Photographs", undertaken by me and supervised by Dr.Salman Khan. Please contact me or Dr.Salman if you need the full paper.

Introduction

With a multitude of applications, there are numerous researches on deep learning-automated image restoration. For experimental purposes, restoration of random old and degraded photos is an interesting problem domain, since they are more difficult to generalize due to the possible variation in the degradation type and strength, and the possibility of compounding degradations (such as noise, scratch/torn, occlusion, low-resolution, etc).

While published image restoration models are usually benchmarked on datasets with similar degradation type, we performed experiments on randomly scraped dataset of old and degraded black-and-white photos (which are then colorized and restored).

Results Summary

  1. Experiments to enhance the colorfulness of known robust colorization training (DeOldify) by fine-tuning the GAN training:

successfully shows the highest colorfulness score:

  1. Experiments to induce multimodality in the produced colors, by implementing this paper on mode-seeking GAN regularization (Mao et al. 2019), unfortunately still requires more fine-tuning:
  1. We then implement some of the latest papers on image restoration with state-of-the-art results, choose the best performing one, and implement a pipeline for an end-to-end restoration:

Results Summary

In summary, our fine tuning result (1) is able to produce more colorful result than baseline model and the latest image colorization models, both qualitatively and quantitatively. Our multimodal colorization (2) is able to produce interesting and diverse colors, but we have yet to be able to fine-tune the model to produce realistic looking result. Our adopted end-to-end pipeline (3) shows the capability to generalize to varying degree of restoration requirements which is present in randomly scraped old images.

Setup

Dependencies

Ubuntu 20.04, Intel Core i7 3770 and RTX 2080 Ti were used to run the experiments.

Dependencies can be installed by:

pip install -r requirements.txt

Download and setup pretrained models

  1. Clone this repository

  2. Download all files from here to your local computer

  3. Move both .pth files to ./models directory

  4. Move FaceEnhancement-checkpoints.zip to ./Face_Enhancement directory, then extract it

  5. Move Global-checkpoints.zip to ./Global directory, then extract it

Prepare dataset

  1. Legacy-Large dataset can be downloaded from here, at Dataset folder.

  2. Alternatively, both training dataset and test dataset (legacy images) can be prepared by:

    • Run image_scraper.ipynb
    • Download ImageNet dataset from here for training
    • Run preprocess.ipynb to resize test images and create BW images

Colorization

  1. Best results for training and fine tuning can be viewed by running inference using Colorizer_GANFineTune_bestmodel.pth. Example inference is shown in colorize_test.ipynb

  2. Results for Mode-Seeking GAN training can be viewed by running the MSGAN_training.ipynb file

Joint with other restoration

The adopted final pipeline consists of preprocessing using Wan et al's (2020) method from here and feeding the result to the colorization model.

To test it, run joint_restoration_test.ipynb

Calculate colorfulness

To calculate colorfulness using Hasler et al's measurement, run calculate_colorfulness.ipynb

Acknowledgement

The final code implemented here are for research purposes only, and heavily based on the implementation from other repositories:

  1. The original deoldify repository (Antic 2019) here, and
  2. The pre-colorization step are from Wan et al's joint restoration (2020) here