Regarding Increase Resolution?
显示 更早的评论
Related to Digital Image Processing:
Person Said: "You are saying we can't increse resolution from one static image"
I asked: Are you talking about interpolation?
Person said: No
I said: You cant increse the resolution of a image from single image, it can possible if you have stack of images capturing with variation of focus plane (Like burst), then you can infuse to make high resolution images, like composite one image.
Person did not reply.
Am I right or wrong?
9 个评论
KALYAN ACHARJYA
2019-5-24
Rik
2019-5-24
I would say that the amount of information you have is limited. You can't increase the number of pixels without decreasing the amount of information per pixel. Even super-sampling methods are best guesses and may visually improve an image, but they are limited by the images they were trained on.
There is a disconnect between percieved detail and mathematical detail, which is why JPEG works (by removing details humans don't see anyway). This deep learning methods attempt to do the reverse: without adding actual information (because it doesn't exist) trying to increase the details for human eyes.
KALYAN ACHARJYA
2019-5-24
KALYAN ACHARJYA
2019-5-24
Rik
2019-5-24
For a pathology slide I wouldn't change the resolution ever. If people want to zoom in, let them zoom in. Don't mess with resolution if you don't want to add something that isn't there.
KALYAN ACHARJYA
2019-5-24
编辑：KALYAN ACHARJYA
2019-5-24
Steven Lord
2019-5-24
Unfortunately, police procedural TV shows (CSI, NCIS, Law & Order, etc.) in particular often give the impression that images have an infinte amount of information. Time wasting warning: this is a link to a relevant page on the TV Tropes website.
Bjorn Gustavsson
2019-5-24
@Steven: There really, really is "no way out"!
(I feel equally clever and ashamed...)
KALYAN ACHARJYA
2019-5-24
采纳的回答
John D'Errico
2019-5-24
2 个投票
I'll take a stab at it.
There are many tmes when we see in the movies, on TV, the idea that you can get almost anything out of an image, expanding a single pixel into a perfectly clear license plate number. But there are limits to what you can see.
Sometimes, we might have a very high resolution image, hundreds of millions of pixels. But when you look at the image on your basic computer monitor, you only see the broad strokes, as if you were standing a several hundred feet away looking at the Mona Lisa. If so, then, yes, you can expand what you are looking at, essentially cropping parts of the image away. This works until you get to the point where the pixels themselves start to be noticeable. But this is NOT an increase of resolution, merely the effect of standing very close to the picture, cropping away the irrelevant parts. And some pictures are taken at a very high resolution indeed.
An increase of resolution can come about IF you have additional information about the image. If the image is assumed to be essentially a smooth surface, then we could use interpolation methods. Those interpolation mathods, IF they can assume that each pixel is smoothly related to its neighbors, then the interpolation can resolve finer details of the surface. (For an RGB image, interpolate each of the red, green, and blue channels independently.) This is not any different from sampling a sine wave at fixed intervals, then using perhaps a spline interpolant to recover the original smooth shape of the sine wave. And this might be acceptable for SOME images, but it is hardly likely to be appropriate for a general image taken essentially as a photograph. There, the image content is hardly likely to be some smooth, well behaved underlying surface. As such, a higher order interpolant that employs a smoothness assumption to recover deeper structrure makes very little sense. You cannot create signal where none exists.
There are other things you might do however. You might be able to recover information in an image in some sense, based perhaps on an assumption of a known motion blur, or perhaps a known camera artifact. This is more of an enhancement issue of course, not a resolution increase.
Also, you can use multiple images, perhaps from a video sequence, using information from all of the images to provide a finer resolution at some location.
So in general, no, from a single image, you cannot simply increase the resolution. Unless of course, you are on TV, where anything is possible.
5 个评论
KALYAN ACHARJYA
2019-5-24
Bjorn Gustavsson
2019-5-27
Kalyan, both Image Analyst and John talks about using multiple images of the same scene, shifted by by a subpixel-width to extract subpixel resolution, this would be some type of slightly generalized deconvolution, to achieve better resolution this way is possible but just as for "normal" deconvolution the image noise is a bit of a problem - since deconvolution is a noise-amplifying operation (it is typically a high-pass-filtering operation and even though noise-reduction methods can be incorporated noise will limit what can be achieved at some point.)
KALYAN ACHARJYA
2019-5-27
Bjorn Gustavsson
2019-5-27
Well, in some sense and ideal case (that is: no noise) you can deconvolve the point-spread-function of the imaging system - to a limit, and it is not unreasonable to argue that is to actually improve the image resolution. But that is typically not by much - in my experience it is possible to deconvolve an image if we know the PSF but only to approximately 1/3 of the 1-D size of the PSF-width, say from a truncated Gaussian with a FWHM of 5 pixels to something like a FWHM of 3 (2) pixels...
KALYAN ACHARJYA
2019-5-27
更多回答（1 个）
Image Analyst
2019-5-24
2 个投票
You "can't" increase the optical/physical resolution in real world units from a single image, though you can increase the digital resolution (number of pixels). It depends on what the person asking means by resolution since it's one of those words that has multiple meanings depending on the context.
There have been attempts at increasing the true, physical spatial resolution. For example Compressed Sensing was a controversial rage several years ago. And the deep learning thing you mentioned. But I think it basically all comes down to making assumptions, though I'm not sure since it's not my field.
However if you make some assumptions you can increase the resolution, for example sophisticated denoisnig algorithms that reduce the noise enough so that you can now resolve two points that before were just a single blurry/noisy mass. For example a wiener filter or other more sophisticated methods.
Of course you can increase true spatial resolution if you have multiple images, like averaging multiple video frames to reduce noise, shifting the sensor in the optical system by a half pixel, denoising the image, etc.
类别
在 帮助中心 和 File Exchange 中查找有关 Image Processing Toolbox 的更多信息
Community Treasure Hunt
Find the treasures in MATLAB Central and discover how the community can help you!
Start Hunting!
Translated by 
