A Framework for Using Custom Features to Colorize Grayscale Images
Loading...
Date
2016-03-30
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
We propose a new framework for automatic colorization of grayscale images by using the composition of features from multiple color images to solve for the most likely coloring with machine learning. The intention of this process is to alleviate the amount of manual labor required to colorize a photograph. Given a target grayscale image and several context color images, our methodology will colorize the target image using the data from the color context images. The algorithm calculates custom user-defined features for each of the context images, and then uses these features to generate a local loss function for the colors at each pixel of the grayscale image. These loss functions are then used to generate a Markov Random Field, which is solved for the most likely coloring using graph cuts. Simple features have been tested such as the luminance or variance of 5x5 pixel neighborhoods, and more complicated features have also been tested involving a combination of luminance values and texture statistics. In general, the we have found that features carrying more texture-based information tend to result in better final colorizations. More complex and descriptive features could be tested, but at the cost of algorithm performance.
Description
Engineering (The Ohio State University Denman Undergraduate Research Forum)