Earlier this month restored images of iconic Hong Kong film stars like Bridget Lin and Joey Wong began circulating on Chinese social media. An AI-enabled touch-up process revealed distinct facial contours and features in old and fuzzy or low-res photos, revisiting the comely visages that made the public fall in love with these ’90s stars. Hashtagged AI-enhanced Lin and Wong pics spawned more than 40,000 discussion threads and 190 million page views.
The WeChat applet You and I that enabled the transformations is a product of Beijing-based Daxie Technology. The online restoration service was launched in December 2018 and by April there were 40,000 people queued to use it, with wait times stretching to days. The company launched a paid version (US$0.15/photo) to help impatient users skip the line.
Traditional image restoration involves filling in missing information and can be handled using a diffusion-based approach that propagates local structure to other areas, or an example method where missing pixels are constructed to maintain consistency with surrounding pixels. These methods however fail when the missing areas are large, and additional input data may be needed by deep neural networks.
Daxie’s AI first learns the difference between high-resolution originals and low-resolution images in a huge dataset. The machine then does its job simply by making up the difference without a reference. The kind of photos that can be repaired depends largely on the deep learning model’s training data and algorithm adjustments for different tasks.
Last July NVIDIA, Aalto University, and MIT researchers presented the paper Noise2Noise: Learning Image Restoration without Clean Data at the ICML conference, pioneering a new deep-learning approach that can remove artifacts, noise and grain; and enhance images by looking at corrupted photos only. The team trained their system using 50,000 images in the ImageNet validation set.
For the 2018 Spring Festival, Tencent Youtu Lab — the image processing, pattern recognition, machine learning, and data mining research arm of Chinese tech giant Tencent Group — launched a free photo-enhancing applet, the AI Multimedia Lab. Users could scan and upload for example their grandparents’ monochrome wedding picture and the Tencent AI would bathe the scene in natural colours.
The Youtu team analyzed millions of public stock photos and trained their algorithm to learn in multiple colorimetric spaces. The team separated images based on Hue-Saturation-Value colour model to help the algorithm identify for example moldy, faded or damaged areas. The team’s internal testing achieved 95 percent recognition accuracy.
For now, the “You and I” applet is mainly optimizing facial restoration in photos. But such AI image enhancement technology also has applications in video repair, animation enhancement, AI medical imaging and treatment and beyond.
Source: Synced China
Localization: Meghan Han | Editor: Michael Sarazen