Thanks.  I'll take a look.  Looks a little like "vaporware", especially since the example (Obama) is certainly fabricated.  The problem with algorithms that claim to get that kind of result is that there simply is not enough data in the original to derive the result.  Hence the fabrication of the result in their example.

Mike

That was my initial reaction.  I believe we've gone about as far as we can go with photographic interpolation algorithms.  Some are a little sharper than the "hybrid" method in Qimage, but generally when you try to shoot for very sharp edges in a "random" photograph like some of the sharper algorithms do, you end up with a photo that looks a little more like a charcoal drawing than a photo.  The reason is that you've broken the relationship between sharpness and image size because there simply isn't enough data in the original to determine whether or not those edges should really be that sharp.  It's a balance, really, but you can't get something for (from) nothing.  The super-sharp edge detection algorithms do work well for non-photographic images like computer graphics (pie charts, screen shots, geometric 3D renderings, etc.).

I also have to say that there's less need for the super-refined interpolation algorithms these days than in the past.  Most people have cameras that have enough pixels to do the job.  The emphasis now is on proper sizing for printing so that you get the most of the pixels you do have, regardless of how many.  Of course, that's what the smart sharpening in Qimage is designed to do.

Mike

Hi Ernst,

The link you gave is interesting, but does not give much detail. However, it appears it is designed to work more with, if you like, information randomly selected, and makes an 'educated' guess at filling in the missing pieces for the rest of it. A bit like weather forecasting in the UK  . Referring to their link on sparcity it states - I have the feeling it is useful where much less data is available, and that data is not spread evenly over the image area, compared to the sort of high density infomation available in our current digital images. It may be useful, however, I suspect, if in some way the image was damaged, by perhaps a great number of the pixels being missed out, on specific types of image, depending where the damaged area was. But that is what the 'clone stamp tool' and others are used for. The example they showed, was sort of reversed engineered in photoshop, afaik, and I do not think the compressed sensing technique would actually work too well on that type of noise. There already exists good techniques for reducing noise in images. I tend to think that the algorithm could end up enhancing the noise, instead of the fundamental image, but as one of the comments mentioned, it may work well in combination with other techniques.

I think the problems which we face with upsizing our images are different. However, for text, at first thought I imagined it managing anti-aliasing, but thinking on it further, it would need limiting somehow, since I could see the whole result being blurred, and there is no chance it could fill in missing letters.  

Many folk seem to be satisfied with repetitive bicubic upsizing. In effect, I suppose that makes its own attempt at filling in missing data, possibly better than doing it with one hit.

Mind you, I am quite used to being wrong

Best wishes,

Ray

A bit more. In the original efforts, the scan, they knew the sort of image they were going to get. The test image - the phantom- was known, too. Both images vaguely similar. The detail they are recovering, is not that high, I think, compared to the detail in one of our landscapes. We will see where it goes.

I looked into those too... a year ago.  They are awful for photographs!  They only work for pixel art (games and graphics with predictable line boundaries).  In photos they produce pronounced artifacts because they try to make connections that don't exist in the more random world of photographs versus the predictable world of computer graphics.

Mike

I just wonder if it isn't a basis for another de-noising algorithm.  Especially when not used as intensely as their sample which is way beyond anything I'd attempt to save.