No more blurry images?

[Axel Hildebrandt]

http://research.microsoft.com/en-us/...imudeblurring/

The complete article and some more information is at the bottom of the page. It would be interesting to know what makes this different compared with existing anti-shake technology but find it interesting that they used a DSLR and not a P&S camera.

[David Stephens]

I would like to see comparative images of the camera's in-camera image stabiliztion and their deblurred images. Those examples seem to be deblurred vs. no IS or deblurring. The talk about depth related blurring makes some since, but these "before" images are all totally blurred.

[Alfred Forns]

Interesting Axel. Anything that will help will be apreciated but not holding my breath

[arash_hazeghi]

Removing the effects of motion blur using various deconvolution algorithms has been investigated throughly in the past 10 years, to my knowledge some cheap P&S and cell phone cameras already use these algorithms. these two references talk about how this is done mathematically.

http://ai.stanford.edu/~schuon/deblu...otion_blur.pdf

http://www.eecs.northwestern.edu/~sd...lur_CVPR09.pdf

In most cases it is very difficult to get a photo that is pixel level sharp with these methods, what is new in the method described here is that they are using gyrosensors and accelerometers to record motion kinetics during image capture, since the kinetics are known this will help with the deconvolution process, but nevertheless you cannot recover all the harmonics that have been filtered by motion blur so it is never going to be like true optical stabilization. Also they need to sample gyrosensors rapidly, cheap gyrosensors are not that accurate and do not settle fast enough. By the time you add high precision electronics to the camera and the peripheral circuitry the cost will equal or exceed that of current lens/sensor stabilizer modules which often have only two gyrosensors, and you get somewhat inferior performance so I don't think it is very promising for High end SLRs and super telephoto lenses. Also note that the sample images are very small, if they show large size it is most likely still blurry at pixel level.

[Axel Hildebrandt]

Thanks, Arash! Would be cool if it were possible to make it work. Wouldn't there also be a problem with in-camera stabilization and IS lenses?

[arash_hazeghi]

Thanks, Arash! Would be cool if it were possible to make it work. Wouldn't there also be a problem with in-camera stabilization and IS lenses?

In principal it is possible to have both systems function at the same time but way complicated because they systems need to talk to each other in order to function in a synchronized manner, as far as I know the current cameras that have sensor-based stabilization like Sony and Olympus do not offer stabilized lenses, or if the lenses are optically stabilized you can only use one system at a time. Generally lens-based stabilization yields superior results and has two main advantages 1) the finder is also stabilized so tracking BIF and keeping AF sensor on the bird is easier (that is why I keep my IS on all the time) 2) the AF sensor also sees a stabilized image which makes it easier to lock, in sensor-based system the AF sensor still sees a shaky image. also the lens needs to project a larger image because sensor is now moving...overall for super telephoto applications the in-lens system is superior and is going to remain so...Canon and Nikon have so far been reluctant to offer sensor-based stabilization, most likely because it offers no benefit over current in-lens IS/VR system...we will see what happens in future models :)

[Axel Hildebrandt]

I didn't know the image in the view finder would be shaky with in-camera IS. I agree that would not be helpful for BIF.

[Emil Martinec]

Removing the effects of motion blur using various deconvolution algorithms has been investigated throughly in the past 10 years, to my knowledge some cheap P&S and cell phone cameras already use these algorithms. these two references talk about how this is done mathematically.

In most cases it is very difficult to get a photo that is pixel level sharp with these methods, what is new in the method described here is that they are using gyrosensors and accelerometers to record motion kinetics during image capture, since the kinetics are known this will help with the deconvolution process, but nevertheless you cannot recover all the harmonics that have been filtered by motion blur so it is never going to be like true optical stabilization. Also they need to sample gyrosensors rapidly, cheap gyrosensors are not that accurate and do not settle fast enough. By the time you add high precision electronics to the camera and the peripheral circuitry the cost will equal or exceed that of current lens/sensor stabilizer modules which often have only two gyrosensors, and you get somewhat inferior performance so I don't think it is very promising for High end SLRs and super telephoto lenses. Also note that the sample images are very small, if they show large size it is most likely still blurry at pixel level.

What Arash said. It seems to me that recording motion data in metadata is a very poor substitute for actual image stabilization. To make an analogy, which would you rather have your camera+lens do: (1) autofocus, or (2) leave the image unfocussed, but record the correct focus distance in metadata for later use by a software deblurring algorithm.

[Roger Clark]

To add to the discussion so far, in any image deconvolution, to recover spatial detail, the cost is increased noise. And that is ignoring artifacts from never perfect algorithms with given never perfect input data. We complain about high ISO noise now. These systems will make it much worse. There is no free lunch unless you are on a CSI TV show ;).

Image motion and focus corrections by software deconvolution have been researched since at least the 1970s when I started getting optical journals.

Arash has it right. In lens IS is superior and is likely to remain so.

Roger

[Axel Hildebrandt]

To add to the discussion so far, in any image deconvolution, to recover spatial detail, the cost is increased noise. And that is ignoring artifacts from never perfect algorithms with given never perfect input data. We complain about high ISO noise now. These systems will make it much worse. There is no free lunch unless you are on a CSI TV show ;).

Image motion and focus corrections by software deconvolution have been researched since at least the 1970s when I started getting optical journals.

Arash has it right. In lens IS is superior and is likely to remain so.

Roger

Thanks, Roger! This is good to know, I don't think they mentioned noise in the paper.

[Emil Martinec]

Any sharpening algorithm has to be able to distinguish noise from pixel-scale image content, so as not to amplify noise as it restores image detail. Deconvolution methods are not exempt from this problem.