Site map/guide to tutorials Contact | News Making fine prints in your digital darkroom Understanding image sharpness and MTF Image galleries / How to purchase prints Photographic technique Image editing with Picture Window Pro A simplified zone system Digital vs. film

View image galleries

Search WWW Search www.normankoren.com

Part 1 describes the four pillars of image quality, digital image sensors, the simulation technique, and presents a summary of results comparing digital camera resolution with film. Part 2 contains Dennis Wilkins' comparison of the Nikon D100 with film, my view of the future of digital cameras, Links, a discussion of Information theory and image quality, and how to measure MTF from Dpreview.com test results. I can't keep up with all the latest camera models. Sites with the latest news and reviews are listed in Digital cameras: Links. Digital camera sharpness measurements are available on Imatest sharpness comparisons.

If you are unfamiliar with MTF, you may want to review Part 1 of this series.

Introduction

---

In September 2002 two full-frame digital cameras were announced: The 11 Megapixel Canon EOS-1Ds and the 14 Megapixel Kodak DCS Pro 14n. These cameras outperform 35mm film, winning the race as much on low noise (absence of grain-- far superior to 35mm) as resolution. The Canon EOS-1Ds Mark II, announced in September 2004, challenges medium format film, and the Canon EOS 5D, announced in August 2005 drops the price of a full-frame 12.8 megapixel SLR to $3,300 USD, not much more than the old 3 megapixel EOS-D30.

This is exciting news because of the many advantages of digital cameras: First of all there is no film, which is the primary factor limiting resolution in the 35mm format. There is very little noise (the digital counterpart of grain) in sensors with pixel spacings larger than about 6 microns. There are no problems with film flatness, inconsistent development, or scanners. You can preview your image immediately-- no trips to the camera shop to drop off film and pick it up, and no worry about airport x-rays.

The four pillars of digital camera image quality-- an overview

---

• Resolution,  by which we mean apparent sharpness and visible detail (closely related to 50% MTF frequency), rather than the pixel count or traditional lp/mm resolution limit. Resolution is the primary concern of this page, but it is only one of several quality factors. For 6 megapixel digital SLRs-- the Canon EOS 10D, the Nikon D100, and the Fuji S2 Pro (which actually has 6 megapixels but uses sophisticated interpolation to simulate 12 megapixels)-- resolution is close, but not quite equal to 35mm film. But overall image quality can be stunning because color quality is excellent and noise-- the digital counterpart of grain-- is nearly absent. For 11+ megapixel digital SLRs, resolution is superior.
• Noise/grain  Noise in digital SLRs is much lower than grain (its counterpart) in film. Digital SLRs exhibit almost no visible noise at ISO 100 and 200, and still have low noise at 400 and 800. Compact digital cameras (with 11 mm diagonal or less sensors) have fairly low noise at ISO 100, though noise can become quite bothersome at ISO 400 and above. Both resolution and noise (actually signal-to-noise ratio) enter into the equation for Shannon's information capacity, which I hypothesize is closely related to perceived image quality.
• Dynamic range  Limitated dynamic range is the principal weakness of digital capture, but it is only a problem in contrasty light, and there are a number of remedies. (See, for example, the article on Tonal quality and dynamic range in digital cameras.) It is most frequency manifested as blown (burned out; flat white) highlights. Film responds to light with a nice "S" curve--as light intensity increases, density gradually levels off in a region known as the response "knee," which is responsible for much of the detail and beauty in the highlights of fine silver-based prints. It is absent in digital sensors, but as long as highlights aren't blown out, prints from digital sources look fine. A "knee" can be created by means of Curves transformations (illustrated here) in image editors. Help is on the way: Fuji's long-delayed SUPER CCD SR technology (announced January 2003, but delayed until at least October) offers a solution using two sensors per pixel (one large, one tiny).
• Color quality  We won't go into detail here, but in my experience digital is consistently outstanding-- better than film when prints are the end product (though not comparable for slides). The reason is that there is less generation loss. With a digital camera, the steps are (1) digital capture, involving filtered sensors, (2) Bayer interpolation (except for Foveon), which doesn't affect color quality, (3) Image editing (no generation loss either, but lots of potential gains), and (4) Printing using pigments or dyes. With film the steps are (1) Exposure, using layered film, where dye-couplers determine the sensitivity of the layers, and (2) development, where the dye couplers  determine the color (light absorption) of each layer. A lot of sophisticated work has gone into making the light sensitivity comparable to the spectral absorption of each layer. If (1) and (2) were the whole story, as it is for slides, film would be ahead, but there's more for prints. For traditional prints you have to (3) expose color paper through an enlarging lens and (4) develop it. This involves the same compromises as (1) and (2). For digital prints you have to go through a process similar to digital camera capture. Bottom line: colors in prints captured on digital cameras are wonderful-- subtle or vivid, as appropriate. They have a three-dimensional quality rarely seen in prints from film originals.

Digital image sensors

---

There are two types of digital sensor: (1) those that use a color filter array (CFA): a filter grid that covers the sensor so that each pixel is sensitive to a single primary color (R, G or B), and (2) sensors that employ the new Foveon X3 technology, where each pixel is sensitive to all three primary colors. Conventional sensors all use CFA's, but Foveon's new technology promises to outperform them if Foveon can deliver.
.
Bayer pattern sensors  Conventional digital sensors have their pixels are arranged as shown on the right, in a pattern known as the Bayer mask (or filter, or mosaic, or CFA for Color Filter Array), with two green pixels for each red and blue pixel. This is appropriate because the eye is most sensitive to green.

With the Bayer mask sensor there is some resolution loss and side effects-- mostly Moiré fringing-- due to interpolation-- the process of filling in the data for the two missing colors at each pixel location. The red and blue pixels are spaced twice as far apart as the green pixels; hence their intrinsic resolution is only half that of green, but sophisticated reconstruction algorithms recover most of the lost resolution as long as some green is present. Interpolation triples the file size (for a given bit depth). For example, in the EOS 10D, a 7 Megapixel RAW file (one pixel per color, bit depth=12, losslessly compressed) is transformed into a 18.9 Megabyte 24-bit file (8 bits per color, which sacrifices some tonal detail) or an 37.8 Megabyte 48-bit file (16 bits per color; which keeps the full tonal detail). (It's always better to edit in 48 bits, but it's OK to store the file in 24 bits once editing is complete.)

You may think of interpolation as the miracle of the multiplication of the pixels. [The Cornell University DSP (Digital Signal Processing) Lab has a strong interest in image interpolation. The highly technical paper, Reconstruction of Color Images from CCD Arrays, by D. D. Muresan, Steve Luke, and T. W. Parks is of particular interest.]

The best interpolation algorithms are iterative-- they are extremely computationally intensive; hence they are best performed on RAW image files in a computer. In-camera interpolation algorithms are less effective and more subject to artifacts such as Moiré.

The 3.4 megapixel Sigma SD9 SLR camera body, reviewed by dpreview.com and photo.net, only takes Sigma lenses. It faces tough competition from 6 megapixel Canon and Nikon bodies, that many photographers can use with their existing lenses. A tough sell but a promising future. [To learn more about the technology, check out US patent 5,965,875-- not easy reading.]

The Fujifilm S2 Pro is a Nikon-mount DSLR with a 6.17 megapixel "Super CCD" that produces up to 12.1 Million (4256x2848) interpolated pixels. Price is $1996 US. (Jauary 2004) I'm not sure how to simulate it. DPReview.com has the URLs of two Japanese sites with sample images: Miscall and Yamada Kumio. It's well worth comparing the ISO 100 images from the EOS D60, D100, and S2 Pro on Kumio's site. Note particularly the block wall just below the clock. The S2 Pro is clearly superior; the S2 images on both sites are amazing. Imaging-resouce and Dpreview.com have published very positive reviews; resoluton is clearly superior the the D60 and D100. The S2 Pro may well equal 35mm. The S2 Pro uses interpolation trickery to double the number of pixels. The idea is that the eye is more sensitive to horizontal and vertical detail than to diagonal detail. Vertical and horizontal detail also tends to be more prevalent in scenes. By arranging pixels diagonally, the horizontal and vertical spacing is reduced by a factor of the square root of 2 (0.7071x). Interpolation takes place not only at pixel locations, but between them, doubling the number of pixels and improving the interpolated vertical and horizontal resolution. The octagonal pixels are said to be a better match for the microlens used to increase the sensor's effective fill factor. Strange, but it seems to work. This excellent post sheds light on how.

Simulation technique

---

The key parameter required to model a digital camera is the sensor's MTF response. Solid data on sensor MTF is extremely hard to find on the Web. This forces us to make assumptions-- to choose models that match test results. As we mentioned in Part 2, the response of scanners and sensors can be approximated by functions of the form |sinc(f/dscan)|ⁿ, where dscan is sensor resolution in pixels/mm.

These functions are simple, have nulls at the correct frequencies, and closely approximate measured data. Their frequency and spatial response is shown in graphs in Part 2 for n = 1 through 4. The spatial sensitivity function for sinc(f/dscan) is a rectangle of length 1/dscan (the length of a sensor element with a fill factor of 1). An "ideal" rectangular sensor with a 100% fill factor would have a sinc(f/dscan) response. For n > 1, the spatial sensitivity is the rectangle convolved with itself (n-1) times. For sinc² it is a triangle of length 2/dscan. As n increases, spatial response spreads.

The MTF of CCD digital sensors is affected by diffusion or "blooming" between pixels. Most digital sensors have anti-aliasing (lowpass) filters, which spread spatial response and reduce high frequency MTF. Bayer sensors have some MTF loss due to their layout and interpolation routines, though sophisticated software keeps this loss to a minumum. Hence no digital camera sensors would be expected to match the sinc(f/dscan) response of an ideal rectangular sensor. After some trial-and-error the following approximations were selected. [I changed the assumptions February 2, 2004 to conform to my experience with the Canon EOS 10D. Digital cameras come out better as a result.]

• The MTF of Bayer sensors with anti-aliasing filters approximates |sinc(f/dscan)|³. Anti-aliasing filters are found in digital SLRs (except for the Kodak DCS Pro 14n). A page on a low cost Kodak anti-aliasing filter illustrates MTF response, which may or may not resemble actual filters used in digital SLRs.
• The MTF of Bayer sensors without anti-aliasing approximates |sinc(f/dscan)|². These include compact digital cameras (with pixel size <= 4 µm) and the Kodak DCS Pro 14n, which has a relatively large 7.9 µm pixel spacing--risky: it could have a problem with Moiré fringing.
• The MTF of a Foveon X3 sensor, which doesn't require interpolation, and apparently doesn't have an anti-aliasing filter, approximates |sinc(f/dscan)|^1.5. The X3 sensor requires less anti-aliasing than Bayer sensors because aliasing is black and white instead of color, hence much less obvious. It works better in situations where green energy is absent.

View image galleries .. .. .. An excellent opportunity to collect high quality photographic prints and support this website .

Here are the results for the 135 pixel/mm Canon EOS 10D (and 300D) with the "excellent" lens, sharpened with ksharp = 0.50 (generated by MTFcurve2 1e4 0 61 2 135 3 .50). The 50% and 10% MTF frequencies are 61 and 84 line pairs/mm, considerably higher than Provia scanned at 4000 dpi and sharpened with ksharp = 0.60 (46.7 and 75.6 lp/mm). There is little aliasing (response above the Nyquist frequency, 67.5 lp/mm). The key result is the black line: the MTF of the lens + sensor + sharpening. The thin magenta curve is the spatial response of the lens only. The kinky red curve is the spatial response of the sharpened digital image. The thin blue line is the MTF of the lens. The blue dashed line is the MTF of the sensor + sharpening. The sensor has a built-in anti-aliasing (low pass) filter. These numbers are consistent with DPReview's resolution chart comparisons. Since the 10D's sensor dimensions are 15.1/24 = 0.63 as large as 35mm, the total resolution at the 50% level is (61/46.7) * (15.1/24) = 0.82 that of 35mm 4000 dpi Provia. That's fairly close, and when you consider the extremely low noise of the digital sensor-- the absence of grain (see the discussion on information theory), overall image quality approaches that of 35mm. Another advantage of the low noise is that you can get away with more sharpening.

The EOS 10D has 82% the total resolution of full frame 35mm Provia scanned at 4000 dpi.
But perceived image quality approaches that of film because of the low noise-- the absence grain.

Digital camera resolution summary

---

Based on the equivalence of 10.2 micron pixels with film, I estimate that a full-frame sensor with 8.3 megapixels would have resolution equal to 35mm film. Slightly more pixels would be required for smaller sensors with lenses designed to cover full-frame 35mm. The simulated resolution of the Canon EOS 10D relative to 35mm film (82%) agrees well with my tests on the 10D. Likewise, the simulated resolution of the Nikon D100 agrees with Dennis Wilkins' tests.

The image on the right was cropped from a 3072x2048 pixel image taken with the EOS 10D. Details can be found on the EOS 10D page. It illustrates how digital cameras make highly efficient use of pixels; images tend to be sharper at the pixel level than 35mm scans. There are no MTF losses from film, film flatness error, development, or scanning (optical and mechanical). Sharpness is consistent and there is little noise/grain. That's why digital camera images achieve comparable perceived image quality with fewer pixels.

In interpreting these results, remember that resolution is not the only factor that influences image quality. Digital cameras with large pixels (over 5 µm) have far less noise (the equivalent of grain) than film, hence they will have better image quality with the same resolution. I find the image quality in my 6.3 megapixel Canon EOS 10D to be equal to 35mm film-- remarkable when you consider that 35mm film exposed and processed with excellent technique (fine lens around optimum aperture, perfect focus, sturdy tripod) can produce remarkably sharp 13x19 inch prints-- finer than most people realize. I haven't used much film (except for the Hasselblad XPan) since I got it.

Michael Reichmann's article, Ultimate Shoot-out, compares his 11 Megapixel EOS-1Ds to the medium format Pentax 67II. Although I agree with his conclusion that the EOS-1Ds has better overall image quality-- I'd buy one today if I could comfortably afford it-- I have to dispute one important detail: which camera has the higher resoluton. It's the Pentax. You can see it on the enlargements of the windows in the middle of the page, the second set of images under The Print Evaluation. A vertical line on the wall on the right is visible in the Pentax image but not in the EOS-1Ds image. Should this change Michael's essential conculsions? No. This thin line represents response at a very high spatial frequency that would have little effect on the appearance of a 13x19 inch image. It would be visible on a 24x30 inch image, but the 1Ds still has superior image quality due to the absence of grain-- see the discussion on Shannon channel capacity and image quality. Sharpness is only a part of the equation. Reichmann's earlier review of the D60, with side-by-side images from the D30, D60, 35mm and medium format cameras, is also worth reading.

The Foveon X3 sensor is an important step above the traditional Bayer mask sensor; images will look better at the same pixel count because of reduced artifacts (Moiré, etc.). But Foveon needs to increase the pixel count of its sensors to compete successfully.

Images and text copyright © 2000-2013 by Norman Koren. Norman Koren lives in Boulder, Colorado, where he worked in developing magnetic recording technology for high capacity data storage systems until 2001. Since 2003 most of his time has been devoted to the development of Imatest. He has been involved with photography since 1964.