Sony A900 Noise Reduction

Sony A900 Noise Reduction

One very encouraging sign in the Sony A900 is the extent to which Sony's engineers have listened to reviewers in their implementation of noise reduction (NR). The Sony A700 was an excellent DSLR, but many (ourselves foremost among them) felt that its Achilles' heel was that its high-ISO noise reduction couldn't be disabled, even in its RAW files.

In our view, a RAW file should be just that: An exact copy of whatever raw data came from the sensor array. Think of it as a "digital negative," which you can then choose to "develop" however you want. In rendering their RAW files, different users may prefer different tone curves, sharpening or noise reduction algorithms, color mapping schemes, etc. The important thing is that these choices be left up to the user.

The problem with the A700 was that its RAW files weren't really RAW: Even if you set the high-ISO noise reduction to "Off," there was always some amount of inter-pixel NR applied before the data was recorded in the RAW file. This meant that, no matter how good the NR algorithms in your post-processing software were, or how much you might want to choose to leave some noise in the image in order to preserve fine subject detail, you had to accept the particular tradeoffs chosen by the Sony engineers in their pre-RAW image processing. Ultimately, the A700's files came to be called "cooked RAW" on many internet forums.

Imaging Resource took the lead in calling attention to this issue in our early analysis of high-ISO images from the A700, so you can imagine how elated we were to hear that, in the Sony A900, when you turn NR Off, it actually does turn off, at least the inter-pixel processing in the RAW files!

Astute readers will note, though, that Sony's marketing materials for the A900 still make prominent mention of NR that's performed before data is saved in the RAW file. Also, in our discussions with Sony product managers and planners, they did say that there is still some NR that's done, even with the NR option set to Off. The good news is that this is "good NR," not the "bad NR" that caused so much foment in the case of the A700.

Wait a minute though: "Good NR"? "Bad NR"? What's that mean? This is an area that could cause a lot of confusion, and hence argument, so we'll take a moment here to explain it.

There are two fundamentally different approaches to reducing noise from an image sensor. On the one hand, you can deterministically measure the contribution certain noise sources are making to a pixel's output and subtract that contribution back out of the final image data. Alternatively, you can reduce the perceptual impact of the noise, through various clever image-processing techniques that act over areas of the image involving multiple pixels.

By definition, deterministic noise reduction techniques are removing only noise from the imager chip's signal, leaving the original image data untouched. This is a good thing, and is why we refer to deterministic NR as "good" NR.

On the other hand, perceptually-based noise reduction relies on characteristics of human vision and tries to remove both noise and subject detail in such a way that we aren't aware of how much subject detail is being lost, or of how much noise is still present. Perceptual NR basically tries to determine how much subject detail is present in a given area of the image, and ratchets the noise reduction up or down accordingly: In areas where there's a lot of local contrast (strong, fine-grained patterns, textures, or edges), the NR algorithm leaves the image more or less alone. There'll be more noise there, but the subject detail will tend to mask it, so it won't bother viewers as much. Alternately, in large areas with very little subject detail present (sky areas, for instance), any noise present will be very visible, so the NR algorithm flattens-out local brightness or color variations.

The problem, of course, is that a perceptual NR algorithm can only guess at what's subject detail and what's noise, so it's inevitably going to lose some of the former in its attempts to eliminate the latter. As the overall noise level increases (at higher ISOs or with smaller pixels), the local contrast from the noise will overwhelm that from fine subject detail to a greater and greater extent. At the same time though, there's a greater need for noise reduction, so the tendency will be to set the threshold for stomping on (flattening out) local contrast lower and lower. The result is the "watercolor" look we've all seen whenever we push the ISO level higher than our camera's sensor is really capable of delivering. Because they unavoidably throw away subject detail with the noise, we call perceptually-based NR systems "bad" NR.

So, back to the Sony A900 and how it differs from previous Sony DSLRs, notably the A700. One of the strengths of CMOS sensors (as used in both cameras) is that the technology allows per-pixel (good) noise reduction. The sensor in the Sony A900 takes advantage of this in all its shots, regardless of the setting of the Noise Reduction setting in its menu system. This is as it should be: Since per-pixel NR loses no subject detail, there's no reason to ever disable it. Per-pixel NR won't get you all the way to where you want to go, though, particularly at high ISOs. The A900 thus applies varying amounts of perceptual-based NR to its images. Unlike most digital SLRs, it applies this to both its RAW and JPEG files. BUT, unlike the A700, when you ask the Sony A900 to turn off its high-ISO NR, it does in fact refrain from applying perceptual NR to its RAW files.

Big kudos to Sony for listening to its customers (and reviewers) in this critical area!