Digital vs film imaging

1
Digital versus film imagingThe issue is channel
capacity
Norman Koren Imatest Boulder,
Colorado www.imatest com

• Film and digital cameras
• Image quality and Shannon information capacity
• Sharpness and tricks for enhancing it
• Noise and tricks for reducing it
• Approximate Shannon capacity as estimate of image
  quality
• The Imatest program for measuring image quality
  will enable educated consumers to test the
  quality of their digital cameras and lenses. Will
  there be a market?
• Results for different cameras (data on film is
  incomplete.)

This is a work in progress. There are still more
questions than answers.
2
Film and digital cameras

• Film cameras my old 35mm SLR (Single-lens
  reflex) is shown.
• Slide film Low grain but low exposure (dynamic)
  range (5 f-stops).
• Negative film High exposure range (9 f-stops
  but more grain.
• Compact digital cameras Small, lightweight. Up
  to 8 megapixels (3-5 typical). Non-interchangeable
  lenses, mostly zooms. Small sensors (?11 mm
  diag.) pixel sites have a limited dynamic range
  and high noise at high ISO speeds.
• Digital SLRs Size and weight comparable to film
  SLRs. 5-14 megapixels (6 typical).
  Interchangeable lenses. Large sensors (?22 mm
  diag.) have low noise. Limited exposure range (6
  f-stops) with standard 8-bit depth JPEG
  conversion, but can have high exposure range
  when RAW conversion takes advantage of the 12
  bit A-to-D converters on the sensors.

3
Three factors driving the evolution from film to
digital photography

• Cost Higher initially than film cameras buyers
  switch when they perceive a savings in their
  individual timeframes.
• Convenience Ease of obtaining the end result,
  which is changing as workflows for obtaining
  digital prints evolve and photographers move from
  prints and slides to electronic display.
• Quality How do the two media compare?
  Quality-conscious photographers will only switch
  when digital quality is perceived as comparable
  to film (various formats). Misconceptions
  persist. The focus of this talk.

Bottom line 300 million digital cameras are
projected to be sold in the next five years.
4
Factors affecting image quality
Important factors, BUT, these can be adjusted in
the camera or during post-processing, providing
the exposure has captured the essential
information.

• Color (hue, saturation)
• Contrast
• Brightness
• Sharpness
• Noise (grain)

Intrinsic factors, related to the information
captured by the camera. If the camera doesnt
capture them, they wont be there (though you can
fake it to a limited degree with sharpening and
noise reduction).
5
Film vs. digitalobservations
Film (negative) Kodacolor 100 Canon 35mm SLR 20mm
f/2.8 4000 dpi scan 5662x3744 pixels Sharp, but
grainy
Digital Canon EOS-10D (6 mpxl) 17-35mm
f/4L 3072x2048 pxls . Nearly as sharp, little
grain
6
Noise, sharpness, and Shannon capacity

• A cameras intrinsic image quality is a function
  of both sharpness and noise (or its equivalent in
  film, grain). Can one devise a figure of merit
  that includes both?
• Hypothesis Image quality is proportional to the
  Shannon information information capacity C, which
  is a function of bandwidth W, noise N, and signal
  S.
• C W log2(1 S/N)
• Some of the complexities are discussed in the
  following slides. These include
• Measuring noise N and bandwidth W.
• How to define signal S.
• The effect of signal processing on N, W, and the
  estimate of Shannon capacity, C.

Looks simple, but its difficult to calculate
the devil is in the details the calculation must
be approximate (for now) thats why Shannon
capacity isnt widely used. Yet.
7
Sharpness
A property of a lens, film, sensor, or system
defined by boundaries between zones
Bar pattern Sharpness defined by 10-90 risetime.
Patterns of increasing spatial frequency (Log
scale)
Sine pattern Sharpness defined by contrast at a
given spatial frequency.
The top half of each pattern is sharp the bottom
is less sharp. How is sharpness measured?
8
Sharpness and spatial..frequency response
Sine and bar patterns are shown with and without
rolloff for a high quality 35mm lens (on a 0.5 mm
virtual target) Amplitude response of bar
pattern Rise distance (10-90) difficult to
calculate for compound systems. The relative
contrast of a sine pattern (pure tone) is
called Spatial Frequency Response (SFR)
or Modulation Transfer Function
(MTF) Multiplicative for compound systems.
Perceived image sharpness strongly correlates
with MTF50, the spatial frequency where contrast
is half its low frequency value. MTF50 is a close
approximation to bandwidth W in Shannon capacity
calculations.
9
Imatest Sharpness calculation derived from
ISO-12233 standard
Results are derived from a slanted-edge image in
a test chart. Algorithm Find average edge
location. Put each line into one of four bins
based on avg. edge. Find mean 4x oversampled
edge, then take Fourier transform of the spatial
derivative.
Upper (black) curve is the average edge. Lower
(black) curve is the Spatial Frequency Response
(SFR or MTF). These results are strongly affected
by sharpening, The dashed red curves are the
edge and MTF response with standardized
sharpening that corrects for oversharpening.
10
Sharpening

• Enhances the perceived sharpness.
• Present in virtually all digital images because
  most look soft without it. Can be applied in the
  camera, RAW converter, and/or image editor.
• Subtracts a fraction of neighboring pixels from
  each pixel. (Radius 2 shown.)
• Boosts contrast at high spatial frequencies
  (boosts MTF50). Amount depends on sharpening
  fraction and radius.
• Different amounts of sharpening in

• different cameras makes comparisons challenging.
• Sharpening enhances edges, but oversharpening
  creates halos at edges. Can look artificial.

Transfer function MTFsharp( f )
(1-ksharpcos(2pifV ))/(1-ksharp) where V
Sharpening radius / pixel spacing
11
Standardized sharpening I

• Different degrees of sharpening makes comparisons
  between cameras challenging.
• Oversharpened images have halos at edges. Can
  look artificial (on right). Common in compact
  digital cameras.
• Undersharpened images do not appear as sharp as
  they could be.

• Standardized sharpening is a strategy to deal
  with the differences in sharpening.
• Sharpen (or de-sharpen) the image with a fixed
  radius (usually R 2 the value used in most
  compact digital cameras) so the MTF response at f
  0.3Nyquist (0.15cycles/pixel) is equal to MTF
  at f 0.
• The response with standardized sharpening to the
  strongly oversharpened image on the right is
  shown by the dashed red ( ) curves.

12
Standardized sharpening II

• This image, from an excellent 11 megapixel DSLR,
  is undersharpened.
• MTF50 without standardized sharpening is 0.264
  cycles/pixel 1426 LW/PH, not as good as the 5
  megapixel compact digital camera on the previous
  page 1459 LW/PH.
• But with standardized sharpening the numbers are
  1872 LW/PH and 1357 LW/PH closer to
  expectations. (Both cameras have excellent
  optics.)
• MTF50 with standardized sharpening is a good
  approximation for bandwidth W in the Shannon
  capacity equation.

13
Noise N (grain in film)
1. No noise 2. 5 added Gaussian noise (pixels
enlarged) 3. 12 added Gaussian noise 4. 3 with
noise-reduction signal processing
RMS noise is the standard deviation (?) of the
pixel levels in a smooth area. Noise increases
with increasing ISO speed for both digital and
film many mechanisms are involved. Digital
cameras increase ISO speed by amplifying the
signal and the noise is amplified along with it.
The visibility of noise depends on its amplitude,
spectral distribution, and on the magnification
of the image. Middle spatial frequencies are
typically the most visible.
14
Noise reduction (cheatin)
3. 12 added Gaussian noise 4. With
noise-reduction signal processing
Noise degrades image quality, but visible noise
can be reduced by means of signal processing,
AKA, CHEATIN. Zone 4 has been smoothed (i.e.,
blurred or lowpass filtered) in low contrast
regions (where adjacent pixels are beneath a
threshold) and sharpened near the edges (where
adjacent pixels are above a threshold). Some
noise is still visible at low spatial
frequencies, where the eye is highly sensitive.
A rapid rolloff of the measured noise spectrum is
strong evidence of noise reduction. The response
on the right is unusual. Typical unfiltered
response tends to be near white (with a slow
rolloff).
15
Noise reduction example
Canon EOS-10D ISO 200(a high quality digital
SLR)
Low frequency noise artifacts are visible in the
sky, which has been enhanced (darkened and
boosted in contrast) for aesthetic purposes. This
is clear evidence of noise reduction in a camera
has an excellent reputation for low noise. The
CMOS sensor evidently requires noise reduction.
16
More on noise reduction
Although noise reduction smoothing often improves
perceived image quality, lost low-contrast detail
can result in a plasticy appearance in lightly
textured areas, like skin. Digital plastic
surgery. Extreme noise reduction is revealed by
the rapid rolloff in the noise spectrum plot,
shown on the right for a 14 megapixel DSLR with a
CMOS sensor. Why cheatin? Because information
capacity C W log2(1 S/N) derived from a
simple targets appears to be increased even
though actual information (low contrast detail)
is lost. Better targets are needed.
The contrast threshold used for noise reduction
reduces the actual Shannon capacity C by the
roughly the same amount as the equivalent amount
of noise N.
17
Signal S the third parameter in Shannon Capacity

• The choice of signal S remains an issue S
  depends on scene contrast, which varies widely.
  Possibilities for calculating C
• Total dynamic range of camera difficult to
  measure requires 16-bit depth (48-bit color)
  files. Better for DSLRs.
• Typical sunny outdoor scene 1601 ratio

• Glossy reflective surfaces up to 1001 ratio.
• Lower signals (101 or less) representative of
  smooth areas like skies.
• Imatest plots C as a function of S. These numbers
  are new and difficult to interpret, so C for S
  100 of reflective target contrast (about 1001
  ratio) is chosen as an artibrary standard.

18
Sharpness and noise summary I
Sharpness Measured by MTF derived from
slanted-edge target. MTF50 correlates with
perceived sharpness. Sharpening enhances MTF.
Some is required present in virtually all
digital images, but the amount varies. Too much
sharpening results in halos. Standardized
sharpening is used for a fair comparison between
cameras. MTF50 with std. sharpening used for W in
Shannon eqn. Noise N Measured by RMS variation
of signal. Amount depends on pixel size, ISO
speed, sensor technology (CMOS vs. CCD),
etc. Visibility depends on spectrum, degree of
enlargement. Noise reduction (NR) reduces N but
removes detail.
19
Sharpness and noise summary II
Noise N N measured with NR can be used in Shannon
equation to (continued) give a number that
correlates to visual quality rather than true
information capacity. Signal S Choice of
signal remains an issue Depends on scene
contrast. Scenes vary widely. Arbitrary choice
use 100 of contrast in reflective target (about
a 1001 ratio). Also plot C vs. S (scene
contrast). Shannon C W log2(1
S/N) capacity A function of bandwidth, signal,
and noise. Difficult to measure exactly an
approximation is required. Because signal
processing masks the true C, we have to settle
for a visual Shannon capacity based on
imperfect measurements.
20
Results 6 megapixel digital SLR
Popular consumer DSLR. Excellent performance for
a wide range of ISO speeds. MTF50(corr) 1370
LW/PH. C (_at_100 of W-B) 4.15 MB.
21
Results 14 megapixel digital SLR
Highest pixel count DSLR (2004). CMOS sensor no
anti-aliasing filter. Extreme noise reduction
apparent in the noise spectrum. MTF50(corr)
3282 LW/PH (higher than total pixel count
aliasing is an issue). C (_at_100 of W-B) 15.5 MB
(somewhat bogus due to NR and aliasing).
22
Results 5 megapixel compact
Popular high quality compact. Significant
oversharpening. Excellent performance at low ISO
speeds. MTF50(corr) 1366 LW/PH. C (_at_100 of
W-B) 2.92 MB.
23
Results 8 megapixel compact
High-end compact. Excellent performance at low
ISO speeds degrades rapidly at high ISO
speeds. MTF50(corr) 1609 LW/PH. C (_at_100 of
W-B) 4.65 MB.
24
Conclusions
An 8 megapixel DSLR has about the same total
MTF50 (in LW/PH) as 35mm film cameras (ISO 100
color slide film scanned at 4000 dpi and
sharpened). BUT since noise is lower (and dynamic
range is somewhat higher) with DSLRs, a 6
megapixel DSLR has comparable image quality (and
Shannon capacity C, though tests are
incomplete). Compact digital cameras have
excellent capacity, though their performance
degrades at high ISO speeds and their dynamic
range is more limited. Shannon capacity is
difficult to measure because W and N are masked
by signal processing and no single value of S
represents all scenes. The Imatest program now
makes it possible for photographers to measure
(approximate) Shannon capacity. Shannon capacity
may well become accepted as a metric for
measuring camera quality when (1) devilish
details in measuring W, N, and S are worked out,
(2) the concepts become more familiar, and (3)
perceptual testing (relating C to perceived image
quality is performed.