Automated Movie Restoration

Trinity College Dublin MAI project -- Prof. Anil Kokaram, Research Xin Shu

Thesis Reference

This repository contains the source codes for the MAI research project specified below. The workflow contains the Aritficial degrade film dataset preparation, CNN model training, and prediction post-processing.

@techreport{AutomateLineScratchDetection,
    author = {Shu, Xin},
    title = {Automated Film Scratch Detection},
    institution = {Trinity College Dublin},
    year = {2022},
    month = {April},
}

Proposed CNN architecture

With this architecture, 3,428,642 trainable parameters and 1,408 non-trainable parameters are involved.

Obtained Resutls

Environment Requirements

The code has been tested with Python 3.6 and TensorFlow 1.15.5 on AMD core device.

conda create --name afsd python=3.6
conda activate afsd
conda install numpy, opencv-python
pip install tensorflow-directml