Super-resolution is the task of producing a higher-resolution image or video from lower-resolution input - in effect, inventing detail that was not present in the original pixels. The “enhance” button from crime dramas is the popular image of it, and while the real thing cannot recover information that was truly never captured, modern methods do remarkably well at predicting plausible high-frequency detail from learned priors about what natural images look like.
Classical super-resolution relied on interpolation or hand-engineered sparse-coding pipelines. The deep-learning era began in 2014 with SRCNN, which trained a convolutional network end-to-end to map low-resolution images to high-resolution ones. In 2016, SRGAN added a generative adversarial network and a perceptual loss so the output would contain believable, sharp texture rather than the smooth, blurry average that accuracy-focused losses produce. There are two flavors worth distinguishing: single-image super-resolution, which works from one frame, and multi-frame super-resolution, which fuses several slightly different captures - the approach behind smartphone “super resolution zoom.”
Super-resolution is one of the most widely deployed pieces of AI in consumer technology, even though most users never hear the term. It powers digital zoom and night modes in phone cameras, upscaling in TVs and streaming, game rendering at lower internal resolution, medical and satellite imaging, and the restoration of old photos and film.
Why business readers should care: super-resolution lets you spend less on capture, storage, or bandwidth and make up the difference with computation. That trade - cheaper sensors and pipes, smarter processing - is a quiet but pervasive source of value in imaging products.