An End-to-End Approach for Automatic and Consistent Colorization of Gray-Scale Videos Using Deep-Learning Techniques

Abstract

Automated colorization of black and white videos has been subject to much research within the image processing and machine learning communities. We are aware of the fact that video colorization, until now, is a computer-based process requiring user intervention at many steps. We present a deep-learning-based process that colorizes black and white videos without any human assistance. The primary objective of this work is proposal of a system that aims to create a fully automated and highly consistent video colorization system. Our primary goal is to demonstrate the viability of our methodology in creating consistently colored videos and reveal promising avenues for future work. Current methods in existence incorporate either frame by frame image colorization which leads to inconsistent results like blurred videos or they use user scribbles which are accurate but are time-consuming and require user intervention. We on the other hand propose a method for colorizing videos using key-frame extraction, colorizing the independent key-frames, and then propagating the colors from key-frames to semantically similar frames.