![]() This perspective certainly has plenty of historical precedent. Was color really an interesting element of a composition, or was it something that could be filled in later as an afterthought by a hallucinating computer program? The idea that so much of color was pre-determined - enough so that a machine could guess at the colors in a scene and get it approximately right - was disturbing and a little depressing. As a professional photographer, color is hugely important, and a big part of my work is getting the colors in an image exactly right, choosing subjects based on their vibrant colors, etc. When I first started testing Colorful Image Colorization on grayscale images, my first react was unease, bordering on disgust. ![]() Original image credit New York Public Library, colorization by Gado with Colorful Image Colorization. The algorithm correctly colored the drink car in this image Coca Cola red, presumably from seeing lots of red Coca Cola logos in its training data. Importantly, the algorithm was never taught what a Coca Cola logo is - through the magic of CNNs, it figured this out from looking at lots of training data. Given a grayscale historical image with a Coca Cola logo, for example, it correctly colors the logo Coca Cola red - presumably from seeing thousands of training images with red Coca Cola logos. Some pretty remarkable emergent properties bubble up in the algorithm’s results. Doing this was a machine from a grayscale original - even 32% of the time - is quite an accomplishment. People in the Turing test didn’t just believe the image they were seeing was a well-executed hand colorization - rather, they believed the image really was a color image. That doesn’t sound like much, but remember, this task was even harder than just plausibly colorizing a historical image. Its creators report that when the results were shown to humans in a “colorization Turing test”, people believed the colors were real 32% of the time. Original photo credit New York Public Library, colorization by Gado via Colorful Image ColorizationĬolorful Image Colorization was trained on over 1 million images. The Colorful Image Colorization algorithm can add plausible colors to black and white photographs. ![]() The algorithm uses several feed-forward passes to ultimately take in a grayscale image, and in the words of the algorithm's creators, “hallucinate” a plausible (though not necessarily correct) set of colors to fill into the image. Training data is easy to obtain here - any color image can be changed to grayscale, and then paired with its color version to make an easy training example. This means you can actually use a Convolutional Neural Network to colorize historical black and white photos.Ĭolorful Image Colorization is an algorithm which uses a CNN to analyze the colors across a set of color images, and their black and white versions. ![]() Skies are usually blue (or could plausibly be blue), greenery is green, people’s skin is skin colored, water is blueish, clothes usually aren’t garish or crazy colors, etc.īecause color is more predictable than you’d think, it’s almost more tractable using Machine Learning than you might initially think. In other cases, though, colors are predictable - surprisingly so. are lost forever the second a black and white photo is taken. In many cases, the colors in an image are unique - the exact color of a person’s clothing, the perfect shade of green for a tree, etc. A Deep Learning ApproachĮnter Convolutional Neural Networks. Even with modern tools, hiring an artist to colorize a single historical photo costs between $300 and $500. You have to make decisions about the colors to add in, have the painting skills to place them into the original photo, etc. ![]() Hand-colored photos are beautiful, but making them is slow work. Hand colored images, like this lithograph of Cincinnati ca 1840s, were works of art. ![]()
0 Comments
Leave a Reply. |
AuthorWrite something about yourself. No need to be fancy, just an overview. ArchivesCategories |