General Optical Low Pass Filter swap project (to reduce moire and aliasing)

That's a good question. What are the D30s active pixel size that the filter is tuned for? If you can get that, we can think about it in terms of what we need.

I am sending another email to the optical engineer to see if I can get some more info for us and maybe even another source of materials.
 
Updated my calculation post with this:

Sensor size = 22.3 x 14.9 mm
18Mp (7D/T2i) Native resolution = 5184 x 3456
18Mp (7D/T2i) Native horizontal pixel pitch = 22.3mm/5184pixels = 4.3um
18Mp (7D/T2i) Native vertical pixel pitch = 14.9mm/3456pixels = 4.3um (hey, square pixels!*)

3Mp (D30) Native resolution = 2160 x 1440
3Mp (D30) Native horizontal pixel pitch = 22.3mm/2160pixels = 10.3um
3Mp (D30) Native vertical pixel pitch = 14.9mm/1440pixels = 10.3um
 
The hard part will be coordinating all of this. The only way to make sure it happens is to get someone involved who wants to make money off of it... it is in our best interest to pay a small premium for this business to make a profit... or it will never happen (ever or in time before Scarlet).
 
This definitely explains why the OLPF filter didn't work. The active area of the D30 pixel is over 2x as large.

The complexity here in using an OLPF is that with a sensor that uses all adjacent pixels, the difference between active area and non-active area is going to be very small. With line skipping, the opposite is true, there are large areas between the active pixels that are being ignored. It makes me feel like trying to use an OLPF is not going to work with line skipping, but rather, using diffusion might work (ie, trying to use the caprock idea behind the lens).

Here is a graphic I made in photoshop to illustrate the problem with using an OLPF for say the D30 on our problem:

OLPF.jpg


You can see there are two different sets of pixel sizes. The second row represent if we had a CMOS array with twice as many pixels and so we are using half as many to get the same resolution. The red line represents a VERY tiny line being projected on to the sensor.

The bottom two images in the first column are what gets recorded at each pixel with no OLPF of any kind. The larger pixels at least pick up the whole line, it's just stair stepped (aliased).

The other sensor has missing parts of the line, it is MUCH worse.

The second column is just a vertical example of OLPF, basically, 2 layers of birefringent that is splitting the line up in the first, then back down in the second. Since the layers are identical, just one rotated 180 degrees, you get two copies of the line that are slightly dimmer than the center.

For the larger pixels, you get dimmer recorded pixels around the brighter center line, just like you see when you zoom in to an anti-aliased line in a program like photoshop (again, remember this is a crude example). On the pixels with smaller active area, this doesn't help at all, it just adds noise.

This is the problem, that's why it's not so simple because all that dead area has detail in it that needs to be distributed out of that area with a filter.

My fear is, because the pixels are so small, the amount of diffusion or double refraction it takes to clean out all the false detail completely is going to make the image look horrible when looking at these issues. Finding a sweet spot where it is reduced more might be a better target, but I think the image quality is still going to suffer.
 
Last edited:
Still no word, from the OLPF company, but I've got another graphic. In this, I took the original thin line and changed the thickness of the birefringent material so it would be the same pitch as the example pixels (roughly) and then added another layer in each opposite direction.

OLPF-2.jpg


The center line should actually be brighter than all the rest due to the overlapping line images, but I am sure you get the point. I then extrapolated the lines on which active sensor areas they touch. As you can see, the image would now be much clearer.

The problem for us is that if we know we skip say two rows (my example skips one) so in each direction, we'd need 6 layers of birefringent since we need to duplicate up and down vertically. then left and right horizontally. that means 12 layers to get the duplicated image to spread out to the pixel size of the sensor. Not sure how thick these are so I am not sure if (a) 12 layers is feasible or (b) what that transmittance is on each layer. After that many layers there might be effects in the color since the graphs I see aren't exactly perfectly flat everywhere either.

I see exactly why they use birefringent material and it's brilliant. With diffusion, you spread out the light more than you need to. With double refraction, you spread out the light just to the other pixels so you are minimizing the reduction in intensity.

I was going to do a test with blur but the blur algorithm when I pushed it hard enough to spread out enough to the other pixel spots made me realize how much light was scattered about that was unnecessary since duplicated images offset is a much better plan.

I'll still try and source some birefringent material if possible to pass along, but I think we aren't going to get what we desire here. Line skipped sensors are just a bad method altogether if you want to avoid any false detail.
 
Well stacking 2 D30 filters is relatively easy for me so long as it fits. Thats 4 layers total, 2 vertically and 2 horizontally but interleaved.
 
I think the problem with that is that the duplicated images of the birefringent layers are going to be too far. Since you can't change the thickness of the layers, you would be pushing the detail to the right pixels.

Look at the first image I posted, you'd be creating the same number of lines as the second image but with the spacing of the first (VERY spread out). Since the D30 filter is tuned for their actual pixel size (not pitch). I think you'd get a softer image but still with noise in it at fine detail.
 
Well I guess the point is that its relatively cheap to test. Also, read somewhere that the birefringent material in the OLPF doesn't just split the image but actually produces a "bokeh" like effect where points become a larger geometric shape with hard-ish edges.
 
Will be interested to know your results, and I agree with the "bokeh" like effect it produces. This is very coincident with the discussion I had with what they could do. It seems they don't always rotate 90 or 180 degrees either so they get all kinds of shapes they can produce with hardish edges.
 
So shao, it sounds like you're going to get a couple OLPFs from the D30 and stack them? That could be very interesting to see the results!
 
Thanks! I am going to be very interested in seeing the multi-OLPF results. My prediction is a much softer image that still has moire and aliasing, just less pronounced. However, I hope that I am wrong! :)
 
No updates yet. Camera is still "out there". However, the double OLPF mod is still planned. No ETA at the moment though.
 
Camera is already at the shop ready for the double OLPF to be installed. Turn around has been estimated at 2 weeks.
 
Well, the experiment is at an end and I'm afraid this solution is not viable. The double OLPF didn't physically fit when they tried to install it. And the single OLPF solution wasn't strong enough. So at least we know now.
 
You must log in or register to reply here.
Back
Top