Lack of Anti-Aliasing Filter Causes More Noise at High ISO's?

[pnoble]

Advertisement (gone after registration)

I was reading Photo District News (PDN) online and found this in an article on the engineering difficulties in making a more compact camera with DSLR performance:

 

Though the M9 was the best of the lot, with such a huge image sensor you'd think it would not even be close. However, like many medium-format digital camera companies, Leica chose not to use an anti-aliasing (or "blur") filter over the sensor. The benefit is more detail and increased sharpness in bright light. The downside is you increase the incidence of noise in high ISO images.....Foveon's Guttosch said. "What the blur filter does is knock down your high frequency contrast. There's a good benefit but there's also a price to pay: more noise in lower light.

 

Am I the only one not to have been aware of this? I thought the noise issue, beyond pixel density, was solely CCD v. CMOS.

[Advertisement]

The Leica Q3 is available for pre-order at B&H Photo Video and Adorama

[carstenw]

Interesting point. It makes so much sense when pointed out. The AA filters in Canon and Nikon cameras have a clear blurring effect when looking at 100% crops, so they must blur information across pixels, in effect reducing the high-frequency content, including noise.

[mjh]

Am I the only one not to have been aware of this?

I for one was not aware of this, maybe because it isn’t so. While you can apply optical means of blurring an image (such as an antialiasing filter), noise – shot noise as well as the noise created in the sensor – will still show up in the image. Noise reduction, on the other hand, will blur the image after the injection of noise and thus destroy any kind of high frequency detail, be it noise, moiré, or actual image detail.

 

The comparison made to the Foveon X3 sensor is quite misleading. It is true that Foveon’s sensor doesn’t have an antialiasing filter, as it doesn’t need one, but the reason for noisy images from that sensor is quite different (and correctly explained in a different quote you’ve left out: “There's a broader spectral sensitivity in the red, green, and blue channels in an X3 sensor than a Bayer sensor and that results in a color matrix that amplifies noise.”).

[carstenw]

I for one was not aware of this, maybe because it isn’t so. While you can apply optical means of blurring an image (such as an antialiasing filter), noise – shot noise as well as the noise created in the sensor – will still show up in the image. Noise reduction, on the other hand, will blur the image after the injection of noise and thus destroy any kind of high frequency detail, be it noise, moiré, or actual image detail.

 

The comparison made to the Foveon X3 sensor is quite misleading. It is true that Foveon’s sensor doesn’t have an antialiasing filter, as it doesn’t need one, but the reason for noisy images from that sensor is quite different (and correctly explained in a different quote you’ve left out: “There's a broader spectral sensitivity in the red, green, and blue channels in an X3 sensor than a Bayer sensor and that results in a color matrix that amplifies noise.”).

 

Okay, you win on the noise coming after the filter part (I was too quick to post there), but the Foveon needs an AA filter almost as much as any other sensor, since like the rest of all sensors ever made, since it samples the world through a pixel grid, and thus is subject to aliasing, according to the Nyquist limit. Only a sensor capable of averaging everything seen in an entire pixel would escape this fate (I forget the term here; fill-factor?), and given the need for electronics and microlenses, such a sensor will probably never be made. The fact that a Foveon sensor samples red, green and blue all in the same pixel doesn't let it escape from this, it just raises the Nyquist limit for this sensor.

[mjh]

but the Foveon needs an AA filter almost as much as any other sensor, since like the rest of all sensors ever made, since it samples the world through a pixel grid, and thus is subject to aliasing, according to the Nyquist limit. Only a sensor capable of averaging everything seen in an entire pixel would escape this fate (I forget the term here; fill-factor?), and given the need for electronics and microlenses, such a sensor will probably never be made.

The fill-factor doesn’t make that much of difference with regard to aliasing. Microlenses take care of redirecting nearly all the incident light onto some photodiode so it is like the fill-factor was approaching 100 percent – though not with respect to the full-well capacity that is critical for dynamic range and low noise.

 

So yes, there could be an aliasing issue in any case, even for a Foveon X3 sensor, in theory anyway. In practice, I am not aware of moiré being an issue with Foveon sensors, mostly thanks to the use of microlenses.

[kalina]

So wait a minute... if this is all true, then it's up to us to use Noise Ninja or something equivalent to fix high ISO images and in so doing the M9's high ISO images post-Noise Ninja will look as blah as high ISO images from DSLRs with AA filters.

 

Overall, it sounds like not having the AA filter is still better than having it.

[pnoble]

Advertisement (gone after registration)

The article states:

 

The benefit is more detail and increased sharpness in bright light. The downside is you increase the incidence of noise in high ISO images.....Foveon's Guttosch said. "What the blur filter does is knock down your high frequency contrast.

This seems to infer that the absence of an AA filter results in higher noise in higher ISO images and seems to suggest that this is a result of high frequency contrast.

 

No? Or is this journalistic (or reader) confusion?

[sandymc]

No? Or is this journalistic (or reader) confusion?

 

Journalistic confusion, or perhaps bad editing after the journalist posted his copy.

 

What I would think that Foveon's Guttosch was actually saying was that signal to noise ratio deteriorates at high frequency if you use a AA filter. Which is true - noise remains the same, but high frequency signals are being attenuated. If you then sharpen in post, noise will increase in those high frequencies. But that is only for high frequencies - sharp edges in the image basically - which are a very small constituent of the image, and only if you sharpen a lot in post. Realistically a minor issue even in theory, and nearly impossible see in practice anyway.

 

Just the Foveon folks trying to put some spin on the horrible noise characteristics of their sensors.....which then got "simplified" to say more or less exactly the opposite to what was intended. IMHO.

 

Sandy

[arthury]

There's a thinner AA filter in the Nikon D3 and D3S compared to the D2X. Noise handling at high ISO is handled just fine in the D3. In the D3S, new firmware and re-design of the sensor made noise even lower without the continual loss of content resolution.