On binning images

[Pere Casals]

I never use binning, other than what the scanner does by design when you select a lower resolution. Which is probably just a variation on simple nearest neigbour binning.

Ted, this is difficult to say, if you take a look inside this datasheet https://www.eureca.de/datasheets/01....G_080131_E.PDF it it's clear that linear sensors have lower resolution modes, this allows for faster scans as less clocks are needed to pull sequential analog values from pixels, and also less A/D conversions are performed. If the scanner uses that feature then this is not a binning, but a lower res acquisition.

Anyway binning in scanning does not require Nearest Neighbor, Edge Sensing, etc algorithms because this is not a Bayer mosaic, for each pixel we have the 3 RGB values after the 3 color rows have flied over every spot. In that situation averaging all pixels in the bin is a perfect procedure.

[Steven Ruttenberg]

The article I posted shows the method of currently binning pixels. When you choose 4x4 for example you using 4 red, 4 blue and 8 green pixels to average over and get 1 super pixel. But as you can see, the bayer pattern complicates the binning process however, the binning is not the issue, the problem is in the demosaicing of the image after binning. The authors show that if you fix the step/s after binning image degradation is minimized.

[Pere Casals]

The article I posted shows the method of currently binning pixels. When you choose 4x4 for example you using 4 red, 4 blue and 8 green pixels to average over and get 1 super pixel. But as you can see, the bayer pattern complicates the binning process however, the binning is not the issue, the problem is in the demosaicing of the image after binning. The authors show that if you fix the step/s after binning image degradation is minimized.

Steven, my view is that this article is related to DSLR area sensors but not to scanner linear sensors "Binning in color image sensors results in superpixel Bayer pattern data, and subsequent demosaicking yields the final, lower resolution, less noisy image."

In Bayer area sensors you read only one color in a photosite, the two other are interpolated.

With scanner linear sensors you don't use that, the sensor "colors" are in 3 (or six) rows, and in each spot you make the 3 RGB readings taking advantage of the sensor displacement that sequentially places R, G and B pixels on the same spot, like it's done when pixel shift feature is used in a A7IIR.

[Steven Ruttenberg]

This may be true, but the concept is the same. It isn't necessarily the binning process, but rather the demosaicing and conversion to the digital image we see. The paper argues that if that is fixed and they provide evidence, then binning can be an alternative to achieve a much less noisy image, but with higher fidelity than if scanned at the equivalent dpi. Also, the natural side effect is a smaller file. Given that, are we still limited then in print size based on the binned dpi to smaller prints than achievable with the non-binned image?

[Jim Michael]

Has anyone tried the approach they use in astrophotography where multiple exposures are combined to remove the random noise?

[Steven Ruttenberg]

I did. Not easy cause the program I used converts things to black and white then you pick bayer scheme to go color again. Itbworks okay on earthly stuff, but not great. Ibhave a friend who takes 7 images digitally then combines in PS with median blend I think. Gets good results. Could do that with scanned images, but PS might puke working on 14 gigs worth of images way I scan.

[Ted Baker]

Steven,

The point that I tried to make earlier which was perhaps lost. Is that when you select anything other the 6400dpi on a V series epson (when using the film guide area) you are doing some kind of binning operation. Perhaps more than 1, i.e. something in the X dimension and something different in the Y dimension.

If you pick something other than 3200dpi, or 1600 then you are certainly doing some sort of software binning as well.

It has been my limited experience that if I use my own combination of resizing and sharpening, I can get a slightly better result, than if I left it to hardware/software combination that epson provides.

But it also takes a lot of extra time, that you need to balance against any very minor benefit.

There are some tools in ImageJ that will allow you to do slanted edge testing. I will get around to measuring the difference I observe between the binning/resize that the hardware provides, against my own combinations, to satisfy my curiosity.

[Pere Casals]

Has anyone tried the approach they use in astrophotography where multiple exposures are combined to remove the random noise?

Not only noise, it also enhances resolution.

then binning can be an alternative to achieve a much less noisy image

Yes, but at the expense of a potential information loss. Denoising algorithms are much more smarter, and way more complex:

[Steven Ruttenberg]

I agree that specific progrsms to denoise are perhaps better, but if one was able ti employ the methods of astrophotography you vould have nirvana and pretty haps I csn try that with a scsnof a black and white image if it works, then I can try a color image. But need to capture many dark frames, bias frames and light frames to characterize the scanner. Once you have these, then I would have to scan the image like ten times make separate scand and save individual filed. Not one scan saved several times. Same is true for the bias, dark and light frames. Would start with smaller files first before trying it on 600mb size files. Need sbout 10-20 frames of each set, so thst is up to 60 frames for the bias, dark and lihht and 20 for the actual image. So you can see that computer power becomes important if this actually works. But will try on files of much smaller size first.

U have a program to do this for astrophotography.

[Pere Casals]

Steven, binning may be recommended when sensor resolution way exceeds optical performance, it is the case of an EPSON, but even in that case it has drawbacks.

Imagine you scan 4800 dpi but in your printer you will need an arbitrary fraction of that, say 2126.62 pixels per inch (in the negative scale), resulting from printer ppi and from the particular crop and print size you want.

The way you resize to 2426.62 it can be important ! Even in the case that the scanner resolved only 2400 effective, it's not the same resizing from 4800 than from binned 2400. In the first case the bicubic algorithms (or other) would have more infomation to perform a "cleaner" interpolation/extrapolation.

A "bit" it's like how we deal with decimals in scientific calculations... or banding effect if editing in 8bits per channel: accuracy not well conserved in the processing chain !

It looks weird that your scanner has x2 (or x3) more hardware pixels (linear) than effective pixels, and that was a heavy burden for antique computers, but this also removes nasty aliasing, Moiré effects and other artifacts that even other expensive gear may suffer. Because of speed, this would be a luxury for a true Pro machine that has to earn a profit dayly.

Depending on the operations you are to do in the digital edition (rotation, sharpening) a binned image may bring on severe issues if requiring all image quality possible for releasing large images.

Let me insist, a sound way to evaluate the digital chain is pasting a 1:1 sized crop of a USAF 1951 scan in a test image that will be edited to be released at max size possible, you will witness what happens in the elements as you resize, rotate, sharpen and compress (if for posting in the web).

Of course there is more than resolving power in sharpness, beyond it the image we release has to sport the precise acutance management for the viewing distance... but the described way evaluates how resolving power is conserved in the chain.


Basicly (IMHO) the oversampling of your epson is an advantage for the digital chain that may compensate other drawbacks that this affordable gear has. An epson can be a very competent tool for LF, but with it we don't press a button to get a sound image like with Pro machines, it requires somebody knowing what he has to do.