![]() “You’re telling me you’re going to make an AI that can reason what it’s looking at in a scene, then look it up in Wikipedia and correctly identify what’s in the image versus what it’s looking up, and it also has to know what the time and place and context of the image is?” Antic asks. The fruits in this image appear to have been colored accurately from a black-and-white original. But Antic doesn’t think that computer programs will be able to interweave historical context in the near future.Ī grocery store, ca. The only way to be sure is to find historical evidence, like a text description or a physical artifact, to confirm the colors in a photograph. The algorithm colored fruits and vegetables accurately - at least, the bananas are yellow and the apples are red - but it’s impossible to actually know if a “recolorized” photo is correct. “When there is uncertainty, such as a person’s shirt color, a colorization system must make a guess.”Īntic recently used his DeOldify program, with added object recognition, to recolor an image of a grocery store from the early 1900s. For example, we know that the sky is typically blue, vegetation is green and the mountains are typically brown,” says Zhang. “For certain regions, we ‘know’ the color with fair certainty. To create more accurate colorizers, computer scientists have bolstered their algorithms with object recognition software. ![]() Many different colors can have the same lightness value, and thus appear as the same shade of gray in a black-and-white picture even if they are very different in reality. That’s part of the reason why recoloring old photos is so hard. Hue, saturation and lightness all contribute to our perception of color, but only lightness is present in grayscale images. “Colorization is fundamentally ambiguous, as we are going from a 1-D signal (grayscale) to a 3-D signal,” built around mixtures of red, green and blue, says Richard Zhang, who developed auto-colorization algorithms - or colorizers - during his doctorate at UC Berkeley. Shades of grey, in the left column, are compared to shades of blue, purple and orange, all of which can appear the same in grayscale. In this image, 9 is the lightest color and 0 is the darkest color. Multiple different colors can have the same lightness value. Despite the large training set, it still makes mistakes, in part because grayscale images lack crucial data that is present in color images. The DeOldify program - marketed as the “ world’s best deep learning technology” for recoloring historical photographs - learned to colorize images by looking at more than a million black-and-white photos that were created from colored originals. “ picking up on whatever rules it can from the data.” “It’s a black box,” says Jason Antic, co-founder of DeOldify. For the computer scientists behind these AI programs, it’s often difficult to understand why these mistakes happen. Computer programs can get a lot of things wrong they have colored white waterfalls a putrid brown and digitally coated the Golden Gate Bridge with white paint. ĭespite the competition, Lloyd probably doesn’t have much to worry about. Lloyd, on the other hand, often spends dozens of hours on each image.Ī photo of New York three ways black and white (left), colorized with DeOldify (middle), and the original photo (right). With names like DeOldify, DeepAI and Algorithmia, they can color a black-and-white photo in just a few seconds. ![]() ![]() Many artificial intelligence tools - computer programs that learn and adapt without human intervention - are taking aim at Lloyd’s profession. “The History Channel showed the same image that had been auto-colorized and it looked terrible.” “I remember for the D-Day Commemoration, I released an image that I colorized of the D-Day landings,” says Lloyd. Computer programs that offer to do the same job don’t meet his high standards. Sometimes, there’s no way to know for sure if he’s found a picture’s true colors.įor Lloyd, who calls himself a visual historian and sells his services to publishers and private clients, getting the colors right in an old photograph is a moral imperative. But finding the right colors to add to a grayscale image is an intensive process, demanding thousands of Google searches, archival records and phone calls with experts. Jordan Lloyd sees historical, black-and-white photos as a blank slate of pixels - a canvas with colorful potential. Can we use AI to predict social unrest? Passant Rabie ![]()
0 Comments
Leave a Reply. |
AuthorWrite something about yourself. No need to be fancy, just an overview. ArchivesCategories |