Video Colorizing with Automatic Reference Image Selection

Abstract

In studies to color black and white videos, it is aimed to make the colored video perceptually meaningful and visually attractive. In this way, old images from the past to the present can be visualized and presented to people in different ways. In this study, the video is colorized by using a model that colorizes images with Convolutional Neural Networks. In this reference-based colorizing study, the frames in the video are colorized according to an automatically determined reference photo. As a new approach different from existing studies, instead of choosing a reference photo for each frame, a system that automatically detects the scene transitions occurring in the video stream and automatically recommends the reference photo suitable for the content of each scene.