Friday, November 20, 2015

Kogakuin University Video Super-Resolution Said to Exceed Nyquist Limit

Nikkei: Seiichi Gohshi, professor of the Department of Information Design, the Faculty of Informatics, Kogakuin University, and Fujitsu jointly develop a new technology employed for the "Xevic" image processing engine of Fujitsu's Arrows NX F-02H smartphone, to be released in late November 2015. Unlike commonly-used "reconstructed super resolution" and "learning super resolution," the super-resolution technology being researched by Gohshi uses an original method called "nonlinear signal processing method." A nonlinear function is used to supplement high-frequency components and reproduce high-resolution components that surpass the "Nyquist frequency (half of a sampling frequency)," Gohshi said.

The 2014 University research report gives a list of recent publications by Seiichi Gohshi:


  1. How is this different from the endless adaptive sharpening algorithms already in use? What makes it unique?

  2. That's great, but is it right? I can produce results higher than Nyquist limit, too, simply by injecting in random data above the Nyquist limit. I would probably want to see a ground truth vs reconstructed and corresponding error at the high frequencies.

  3. Their papers talk about creating above-Nyquist frequencies by making the edges steeper, and proceed by calling it "super-resolution", based on a naive idea that additional spatial frequencies equal higher image resolution. Sharpening is something that has been used in all cameras and TV sets for 60+ years. A sharpened "vertical" edge has almost infinite frequency spectrum, for sure, but image resolution is not improved, only acutance. Acutance can be changed post-capture, resolution is a much different story related to sampling in the image sensor. In contrast, there are methods of computational super-resolution that require heavy-duty image processing to extract additional image detail based on a variety of image priors, such as sparsity in certain mathematical domains, non-local self-similarity, simlarity across different image scales, and other image priors. The very fact that their publications never provide any objective measure of their "super-resolution" enhancement, such as PSNR relative to a ground truth high-resolution image, is peculiar. This kind of enhancement counts on subjective perception of increased sharpness while never increasing the actual image resolution. At best this is a misuse of the established terminology, or a plain misrepresentation.

    1. That's what I was getting at with my previous comment, and the buzz in the title makes it sound like they've devised some magic to pull information that is lost out of thin air. Huge red herring.


All comments are moderated to avoid spam and personal attacks.