"Digital Intermediate" where film is scanned, digitally manipulated, then output back to film, offers immense flexibility in manipulating and enhancing the image, including sharpening and grain reduction:
Digital Intermediate
Kodak's Glenn Kennel Talks About Digital Intermediate
O Brother Where Art Thou?
Kodak and the Future

------------------
John P. Pytlak, Senior Technical Specialist
Worldwide Technical Services, Entertainment Imaging
Eastman Kodak Company
Research Labs, Building 69, Room 7419
Rochester, New York, 14650-1922 USA
Tel: 716-477-5325 Cell: 716-781-4036 Fax: 716-722-7243
E-Mail: john.pytlak@kodak.com
Web site: http://www.kodak.com/go/motion

quote:

---

As early as the sixties, NASA reported on a precise technique to refocus poorly focused images based on weighted averaging. Some of the current imaging programs use a popularized version of that technique in their "sharpen" filters.

---

Is this a convolution technique that you're referring to, or another, more complex technique?

Most sharpening filters in today's software are implemented as convolution masks, which is a weighted average of points within a square neighborhood of each given point in the image. Believe it or not, if you're willing to make the convolution mask large enough, it is possible to implement nearly any frequency-domain filter in spatial domain by finding a convolution mask using a least-squares approach which approximates the frequency function. The main reason that NASA is interested in doing such processing in spatial domain rather than in frequency domain is that convolution can be done as new scan lines of the image data become available for processing, and it's straightforward to implement, although for large images and convolution masks, can get very computationally expensive. Try convolving a 1024 by 1024 image with a 1024 by 1024 convolution mask without using FFTs and IFFTS to convert to a frequency-domain point-by-point multiplication of the transforms. Normally, small convolution masks, such as 3x3, 5x5, 7x7, are used, which don't create such a computational burden, and small masks can give surprisingly good blurring or sharpening results.

About the thresholding you mentioned on the image averaging topic: The more knowledge you have about the actual noise, the better off you are in being able to improve the algorithm. If you know that the noise is not random all of the time, and "dropouts" or "spikes" occur in small parts of the image, then throwing out any point outside a certain range of the median of a point's neighborhood, or the median of the point's value from all of the images, will result in a better average. However, this improvement assumes more knowledge about the nature of the noise than assuming it is random. If the noise is truly random and you have absolutely no other information about it, then you're just as well off doing a straight average. If you can determine a good approximation of the standard deviation of the noise, then you have a good basis for determining a reasonable threshold to use in rejecting bad points.

Common sharpening filters are edge enhancement oriented, which is a weak form of enhancement.

From Photoshop help:
"The Unsharp Mask filter locates every two adjacent pixels with a difference in brightness values that you specify, and then increases the pixels’ contrast by an amount that you specify. In addition, you specify the number of surrounding pixels to which the sharpening effect is applied."

Now to describe the NASA system, as I remember it. I couldn't find the old paper abstract, and I am not a math whiz, so I'll try to describe the concept in lay terms.

The first part was a recognition that each lens is different and it was ideal if the original lens could be on hand to examine it's individual characteristics. However, the general concept could be applied to any lens that had an out-of-focus image.

The concept relied on the out-of-focus image being more or less on a single depth plane. It took the known circle of confusion for a specular and used that as a template for a circle, where if there was evidence of a similar circle on the image, that circle would be negated and a corresponding point source inserted at the center. That point became a wild card in further image extrapolations.

You can visualize one of the patterns involved by viewing the airy disk found in a defocused telescope. A star is for all intents a near perfect point source of light, so the rings found when it is defocused in a telescope gives an indicator of exactly how that lens system will deform any defocused image.
Star Test

Next, you can see a diagram of how an image that is defocused will subtend a larger area on the negative.
Auto focus example The important points to remember are that the defocused point image will be circular and there will be a pattern to the brightness in the circle, generally with the center being brightest, with a tapering off to the edge, where the diffraction pattern will form "ripples."

The weighting to negate the defocused bright spot is actually close to an inverse of the diffraction chart shown on this site Lens Tutorial

As I remember, the idea was to pick the peak speculars first, correct them, then notch down a level and correct the next brightest speculars and so on, each time checking for the presense of the characteristic pattern. The process continued to the limits of the resolution and gamma curve of the film.

I see all the math on the sites, but I'll confess it is beyond me and I have little interest in massaging the figures to verify the concept. I have no doubt that it would work on a star field, and I suspect it would work quite well on high contrast subjects photographed out of focus on high resolution film. It would work poorly on low res. digital images.

Finally, as I remember, there was the idea of having different patterns for motion smearing, that would do much the same thing. I'm guessing that image enhancement techniques like this were used on the first Hubble images.

So much for that.

I disagree with you about averaging all points in truly random noise. Intuitively, it doesn't make sense. First, true randomness is surprisingly rare, and "random" noise may not be entirely random, but could follow a curve, making it possible as you suggest to toss out values outside of the median. In any case, the image itself is not random, and greater weight of importance needs to be given to values falling within the expected values of the image. I believe this is where the reduced contrast you were talking about creeps in to your enhanced images.

I'm not sure how the 'digital intermediate' as described falls within the center scope of this, although it is an interesting sidebar.

On the light side... When I see the "digital enhancement" of images of UFOs, I have two reactions. The first is to laugh out loud, since the enhancement is usually a cheesy and blocky false color image, and the second is to wonder just how many of these could be properly resolved by the use of a combination of the technique I just described and multi-image averaging. I suspect that a lot of planes and balloons and fakes could be found.

Whatever. I don't think I'm giving away any state secrets, since what I have related was in an early open NASA publication, and the technical expertise to pull this all off would have to be substantial.

 |  IP: Logged

When I access at home on a dialup through a different ISP, it's all good.

Brad, next time I have a delay problem, I'll email you the time and my IP address so you can check the request logs.

Now back to our regularly scheduled topic ...

 |  IP: Logged

I'm not discounting the possibility of the DNS routing, but I've yet to experience it.

No problems today, but here's today's stats for future reference:

code:

---

Pinging www.film-tech.com  [207.168.10.69] with 32 bytes of data:
Reply from 207.168.10.69: bytes=32 time=94ms TTL=241
Reply from 207.168.10.69: bytes=32 time=95ms TTL=241
Reply from 207.168.10.69: bytes=32 time=95ms TTL=241
Request timed out.

---

code:

---

Tracing route to www.film-tech.com  [207.168.10.69]
over a maximum of 30 hops:
  1     3 ms     2 ms     3 ms  10.1.1.254 
  2     1 ms     1 ms     1 ms  10.100.1.252 
  3     3 ms     2 ms     2 ms  206.141.185.1 
  4     7 ms     7 ms     7 ms  199.179.99.1 
  5    10 ms     9 ms     9 ms  199.179.0.242 
  6    17 ms    17 ms    17 ms  atm3-0x6.nap1.mich.net [198.108.22.230] 
  7     *        *       17 ms  acr2-serial6-1-0-0.chicagochd.cw.net [208.172.1.169] 
  8    18 ms    18 ms    18 ms  corerouter2.willowsprings.cw.net [204.70.9.146] 
  9    20 ms    18 ms     *     core3.willowsprings.cw.net [204.70.4.25] 
 10    19 ms    19 ms    21 ms  atm1-0.chi-c100.gw.epoch.net [206.220.243.61] 
 11    73 ms     *       72 ms  pos0-1.sfo-c000.gw.epoch.net [155.229.123.141] 
 12    74 ms    73 ms    74 ms  pos8-0.pao-c001.gw.epoch.net [155.229.57.197] 
 13    81 ms    80 ms    81 ms  pos3-0-0.lax-c100.gw.epoch.net [155.229.120.50] 
 14    90 ms   112 ms    90 ms  pos0-0-0.dal-m100.gw.epoch.net [155.229.123.126] 
 15    95 ms    95 ms    95 ms  209-101-238-138.dal-m100.cust.gw.epoch.net [209.101.238.138] 
 16     *       96 ms     *     www.film-tech.com  [207.168.10.69] 
 17    97 ms    96 ms    98 ms  www.film-tech.com  [207.168.10.69] 
Trace complete.

---

That's kinda neat to see ... Detroit -> Chicago -> San Francisco -> Los Angeles -> Dallas

 |  IP: Logged

 |  IP: Logged

I hate cache. Sometimes it is necessary, but most of the time I wish that it didn't exist. Without it is the only way to guarantee that everyone is getting the latest updates on each page. And proxy cache is even worse! Oh, and I just loooove the proxies that don't update themselves but once a day!

 |  IP: Logged

 |  IP: Logged