OLPF's

[Wayne Morellini]

Taken some time out to look at that sight Pawel. Its nice but there are some presumptions, people don't know they are presumptions because they are assumed as fact, the compete facts. Those things are overcomable. It is a matter not to be subject to your environment but to master it. To engineer the solution. Seen it many times in engineering, accepted fact. Reading those articles, I realised my alternative lens and sensor design technology would answer a lot of these long-term camera issues, not just price, quality and compactness issues.

http://imagealgorithmics.com/RGB_CFAs_Tested.html#RandomRGB

Charles, interesting. I have given no space to random rgb sensor filters, because a true random at larger scales is going to have concentrations of pixel that produce artifacts on the same part of the screen. But I see this is not at all random, but rather a shared green border with alternating red/blue/green/red and blue/red/green/blue layered 45% rotated box pattern array, or the alternatively offset layered cross pattern description. Have thought of a number of differing patterns in the past, but ussually don't consider one so wide, at the most a zigzag RGB stripe.

I know this is a bad example, as it compares a more gappy foveon, but something I could see before, is that the aliasing contains details, muffled but still there. The olpf camera contains less details on average, a bland "paradise". But the image is too low Res for a proper comparison, even if it was a better d15. But when you think of it, how does Bayer recreate the missing color channels per pixel reliably, couldn't such a thing be applied differently to estimating obscured detail in aliasing (not what I was thinking of doing but a simplistic option that snapped into my mind).

 

[Wayne Morellini]

I forgot, going back to the 100% fill factor, I can see the rare case where you would get strobing stripes from a boxed down sample with an misaligned progressive phase pattern close to the pixel frequency that alternatively alternating the pixels up and down across a spatial domain. However, the solution is the same.

 

[Pawel Achtel]

no OLPF

 

[Christian Edwards]

i don't know what's worse... the aliasing or that suit ?!

 

[Les Dittert]

Pawel, that is unprocessed footage with no OLPF

 

[Charles_Z]

Pawel, that is unprocessed footage with no OLPF

Or it could be processed but with a horrible image scaling kernel.

 

[Wayne Morellini]

no OLPF

Can I buy a suit like that

People don't understand the problem space, all they see are problems, not solutions. Like looking at a dirt road and not seeing how it could become a bitumin road because it is currently a dirt one.

 

[Dan Hudgins]

Worse than aliasing is chroma moire because the red and blue pixels are on every other scan line.

You cannot get color for each pixel in a Bayer sensor without using an OLPF because colored light cannot just fall on one pixel, if you have a red point of light falling on a blue or green pixel because the Bayer filters are very de-saturated you will get JUST a blue or green signal even with only monochromatic red light hitting the sensor.

Chroma artifacts in the circle of confusion make for small rainbows over the pixels, without an OLPF you will get false edge color artifacts combined with chroma moire.

The concept that using a poor quality lens "fixes" the aliasing issue "better" than using an OLPF is very wrong headed, the point of the OLPF is to block high frequendy image details from getting to the sensor at all. Even a bad lens out of focus has sharp edges to the circle of confusion for highlights and light points that are out of focus and those circle of confusion's hard edges and rainbow fringes etc. make up high frequency detail on the sensor.

Also using the OLPF makes the data more compressable since it reduces edge contrast. I have been involved with the development of some digital cinema cameras and have seen the results with and without filters and using various quality lenses and there is no doubt that the results are better overall with the OLPF since it blurs the choma while allowing the luma to be taked from all pixels mixed and then sharpened. You should avoid sharpening the chroma in the de-Bayer processing since that in some ways counters the whole point of the OLPF to make the chroma signal represent the color of the subject even for edges and small points and surface texture.

On the point of needing an OLPF even with 100% fill factor sensors, or CCD etc., yes you always need an OLPF filter, because CFA do not resolve chroma of the subject on each pixel without the small area blur to alow for extraction of the mixed signals of adjacent pixels.

In addition to that, ALL digital images cannot have sharp edges down to the pixel without aliasing, that is why when you generate computer generated images you must apply edge blur as part of the anti-aliasing processing, by oversample or other methods.

Even an monochrome camera with high fill factor needs an OLPF, you can download monochrome DNG from the Acam dII (tm) web site and sharpen those to see what happens with the aliasing on the hairs of the faces shots.

Additionaly having the OLPF allows better control over the luma sharpen and gives more descrepency between the chroma and luma edge issues when doing the de-Bayer processing, because you can not sharpen the chroma and you can sharpen the luma, and in so doing not make the chroma moire worse through sharpening both or having both at some compromise sharpness by default set by the lens focus.

In the point of the DOF adapter and not using an OLPF, in that case focus errors in the alignment of the DOF adapter are acting as a kind of OLPF setting the minimum circle of confusion.

 

[Wayne Morellini]

Worse than aliasing is chroma moire because the red and blue pixels are on every other scan line.

You cannot get color for each pixel in a Bayer sensor without using an OLPF because colored light cannot just fall on one pixel, if you have a red point of light falling on a blue or green pixel because the Bayer filters are very de-saturated you will get JUST a blue or green signal even with only monochromatic red light hitting the sensor.

Chroma artifacts in the circle of confusion make for small rainbows over the pixels, without an OLPF you will get false edge color artifacts combined with chroma moire.

The concept that using a poor quality lens "fixes" the aliasing issue "better" than using an OLPF is very wrong headed, the point of the OLPF is to block high frequendy image details from getting to the sensor at all. Even a bad lens out of focus has sharp edges to the circle of confusion for highlights and light points that are out of focus and those circle of confusion's hard edges and rainbow fringes etc. make up high frequency detail on the sensor.

Also using the OLPF makes the data more compressable since it reduces edge contrast. I have been involved with the development of some digital cinema cameras and have seen the results with and without filters and using various quality lenses and there is no doubt that the results are better overall with the OLPF since it blurs the choma while allowing the luma to be taked from all pixels mixed and then sharpened. You should avoid sharpening the chroma in the de-Bayer processing since that in some ways counters the whole point of the OLPF to make the chroma signal represent the color of the subject even for edges and small points and surface texture.

On the point of needing an OLPF even with 100% fill factor sensors, or CCD etc., yes you always need an OLPF filter, because CFA do not resolve chroma of the subject on each pixel without the small area blur to alow for extraction of the mixed signals of adjacent pixels.

In addition to that, ALL digital images cannot have sharp edges down to the pixel without aliasing, that is why when you generate computer generated images you must apply edge blur as part of the anti-aliasing processing, by oversample or other methods.

Even an monochrome camera with high fill factor needs an OLPF, you can download monochrome DNG from the Acam dII (tm) web site and sharpen those to see what happens with the aliasing on the hairs of the faces shots.

Additionaly having the OLPF allows better control over the luma sharpen and gives more descrepency between the chroma and luma edge issues when doing the de-Bayer processing, because you can not sharpen the chroma and you can sharpen the luma, and in so doing not make the chroma moire worse through sharpening both or having both at some compromise sharpness by default set by the lens focus.

In the point of the DOF adapter and not using an OLPF, in that case focus errors in the alignment of the DOF adapter are acting as a kind of OLPF setting the minimum circle of confusion.

A very sensible post Dan, the most sensible reply post I've had.

You may remember back in the begening of these cheap Cinema cameras, I was no fan of Bayer because of the way the chroma sampling missed a lot of pixels. But they had ways to calculate out these things, well Arri did. I saw an article on it, and judiouse use of debayering routines did the rest. However, I refer to a perfect pixel, as in a 3 color pixel as the base of discussion. You could still get rid of the chroma problems in Bayer in processing, but it is more difficult, and it is possible to miss it at times. But thinking about it, maybe that can to, most of the time. However, going back to present techniques Dan. Back then Arri was detecting different colors through their flow through adjacent pixels, leaving a minor signal in other channels from the cross over response between the color filter ranges. As colors pool, chroma tend to be consistant acoss areas of the color, so the color could be assumed from the surrounding pixels. I wonder if the past Arri choices were part of some sort of superior method. Anyway, I advocated originally, that the presumed color could be a intermediate of adjscent pixels of that channel. Going a step further, giving the slope of the image intensity in the channel, and this color I think, over area, you could guess and adjust the intermediate more accurately through analysis. Now when the color ended, you could work out the border position more accurately. Using the Arri method (thinking about it, it might be another firm) with this, you could get a fairly good, 4:4:4 reproduction. However, now you have to calculate out the negative aspects of the olpf. But the methods I described lead to more natural sharpening, which might describe wat I see in Arri footge, though they may lack my method, they have their own. So a camera test on various aliasing/mioring patterns should be able to tell. Another thing that helps, is to know the characteristics of the sensor structures, and optical path as well, to correct things more accurately, and debayer better as part of this. Anyway, what this 4:4:4, reproduction means is back then I realised that Bayer could be used to store 4:4:4, as a codec, offering 3:1 smaller size, and that the Bayer itself could be compressed somewhere between 3-6:1 loss less, offering 10-18:1 Losless 4:4:4 compression. Not bad for somebody that is not supposed to know anything. Anyway, to get it right you have to store also the difference between the original and the Bayer reproduction at compression time, and use identical debayering, plus the compressed difference on the decompression, to realise the original 4:4:4 signal. So you are going to loss some of that 10:1-18:1 likely, but if you do a very good job at it, mostly not much. I wanted to use the cineform raw Bayer codec with it to get fat 36mb/s TV channels of at least visually Losless (maybe the visually lossless was 18mb/s, I forget). Adding so inter frame compression techniques, you should get that lossless down a lot more. There were a couple more techniques to all this as well. I reserve all the rights of technical publication period. What do you think Dan?

Anyway, onto additional better techniques than these now. I remember in the 90's a top engineer arguing with me about my compressiin philosophy behind this, because something he was taught from a text book at university said it was not possible, but I had, in effect, not knwung the law, found an exclusion aound it, the way law was constructed was a bit niave. Same as I had the alaising, based on other 90's techniques I had come up with.



I've discovered researchers are progressing onto my sensor design, so the future is coming and it is not here yet. Maybe Red will adopt.

 

[Dan Hudgins]

Cross connecting pixels before doing the de-Bayer processing can cause issues with mixing various noise sources between adjacent pixels, such as FPN.

If the OLPF is used to "cross connect" the signal between pixels, the noise sources are specific to each pixel rather than contaminating the recorded information of other pixels.

 

[Wayne Morellini]

Dan, you have totally mistaken what I said, unless you are talking about a new subject. You also seem to be missing the latter part of the first paragraph, which is significant.