background image
Steps towards 'undigital' intelligent image processing: Real-valued image coding
of photoquantimetric pictures into the JLM file format for the compression of
Portable Lightspace Maps
Steve Mann
, Corey Manders
, Billal Belmellat
, Mohit Kansal
and Daniel Chen
EyeTap Personal Imaging (ePi) Lab
Dept. of Electrical and Computer Engineering, University of Toronto
10 Kings College Rd, Toronto, ON, Canada M5S 3G4
E-mail: [mann, corey, billal, mohit] @
Human Media Lab (HML)
School of Computing and Information Science, Queen's University
Kingston, ON, Canada K7L 3N6
Recent advances in intelligent signal processing have
made it possible to capture high dynamic range im-
ages which are better represented as an array of real
numbers rather than the current convention of an ar-
ray of integers. This paper proposes a solution to ad-
dress the need for real, rather than just integer, im-
age coding and file formats. Additionally, we propose
that the real-valued data be linear in photoquantity
(the quantity of light received by the camera) to avoid
the image misrepresentation that occurs when a cam-
era's non-linear dynamic range compressor and a dis-
play's dynamic range non-linear expander do not match.
We present two novel image formats that achieve this:
the Portable Lightspace Map (PLM) and its compressed
version the JPEG Lightspace Map (JLM), that builds
upon the JPEG compression scheme. The results of
various compression levels for real-valued data and their
corresponding file sizes are reported.
1. Introduction
The explosive growth of the digital cameras in the
consumer electronics industry has caused an increase in
spatial resolution by a factor of ten linearly (approxi-
mately 100 times the number of pixels) over the past
ten years, and spatial resolution increases roughly ex-
ponentially with time, following roughly Moore's law,
e.g. roughly 100 times the number of pixels every ten
years. The digital camera industry and consequently the
general consumer, has been focused on the "megapixel"
race, or spatial resolution, while the importance of tonal
resolution has been largely ignored. In fact, over the
years, tonal resolution has remained relatively constant,
at one byte (256 levels), or at most, 12 bits per pixel
per color channel. Existing file formats for image en-
coding are capable of only this fixed tonal resolution,
which means that much of the tonal information in a
scene is lost.
In this paper we discuss the need for real valued rep-
resentation as it applies to image formats. In addition,
to preserve the tonal representation of an image, we
propose that the real-valued data be linear in pho-
toquantity (the quantity of light received by the cam-
era) to avoid the misrepresentation of an image that
occurs when a camera's non-linear compressor and a
display's non-linear expander do not match. Our goal
is to achieve an 'undigital' image representation which
is independent of the capturing or displaying medium's
particular properties, and also minimizes greatly quan-
tization error through real valued representation. Our
solution is presented through the two novel file formats,
the Portable Lightspace Map (PLM) and its compressed
version, the JPEG Lightspace Map.
2. Image range compression
Most cameras do not provide an output that varies
linearly with light input. Instead, most cameras con-
tain a unique non-linear dynamic range compressor, as
illustrated in Fig. 1 which varies widely in its response
function according to the particular camera system.
Historically, the dynamic range compressor in video
cameras arose because it was found that televisions did
not produce a linear response to a video signal. In
particular, it was found that early cathode ray screens
provided a light output approximately equal to voltage
raised to the exponent of 2.5. Rather than build a cir-
cuit into every television to compensate for this non-
linearity, a partial compensation (exponent of 1/2.22)
was introduced into the television camera at much lesser
cost since there were far more televisions than television
cameras in those days.
Coincidentally, the logarithmic response of human vi-
sual perception is approximately the same as the inverse
of the response of a television tube (e.g. human visual
response turns out to be approximately the same as the
response of the television camera) [4]. For this reason,
processing done on typical video signals will be on a
perceptually relevant tone scale. Moreover, any quanti-
zation on such a video signal (e.g. quantization into 8
bits) will be close to ideal in the sense that each step
of the quantizer will have associated with it a roughly
equal perceptual change in perceptual units.
Most still cameras also provide built-in dynamic
range compression. For example, the Nikon D2h cam-
era captures internally in 12 bits (per pixel per color)
and then applies dynamic range compression, and finally
background image
Figure 1. Typical camera and display : Light from
subject matter passes through lens (typically ap-
proximated by simple algebraic projective geome-
try, e.g. an idealized "pinhole") and is quantified
in units "q" by a sensor array where noise n
also added, to produce an output which is com-
pressed in dynamic range by a typically unknown
function f . Further noise n
is introduced by the
camera electronics, including quantization noise if
the camera is a digital camera and compression
noise if the camera produces a compressed out-
put such as a jpeg image, giving rise to an output
image f
(x, y). The apparatus that converts light
rays into f
(x, y) is labelled CAMERA. The im-
age f
is transmitted or recorded and played back
into a DISPLAY system where the dynamic range
is expanded again. Most cathode ray tubes exhibit
a nonlinear response to voltage, and this nonlin-
ear response is the expander. The block labelled
"expander" is therefore not usually a separate de-
vice. Typical print media also exhibit a nonlinear
response that embodies an implicit "expander".
outputs the range­compressed images in 8 bits (per pixel
per color).
Some of the newer cameras, such as the Nikon D2h
camera also allow output of images in a non­range­
compressed "CCD-RAW" 12-bit (per pixel per color)
format. However such CCD-RAW outputs are typically
unstandardised and proprietary.
2. 1 Range compressor expander mismatch
When range compressors were built into video cam-
eras for the purpose of capturing data to be reproduced
remotely, the display devices were all televisions with
largely the same response to a video signal. Today, one
may capture images with no notion of what the image
may eventually be displayed on. The archived images
may be displayed on analog or digital televisions, video
projectors, electric eyeglasses (such as eyetap devices),
print, volumetric displays, etc., just to name a few of
the current possibilities.
In the future, the number
of options will certainly grow. Most importantly, the
range expander in each of these devices will most likely
vary. This means that to accurately display the image,
with correct tonal representation, careful calibration is
needed. In many cases, the calibration may be useless.
This is simply because the range compression used by
particular camera companies is proprietary, prohibiting
accurate representation by arbitrary display devices. In
essence, range compression results in a large probability
of the compressor and expander not matching, resulting
in improper tonal representation of images.
2. 2 Range compression made one byte integers
One of the reasons for the slow progress of any stan-
dardized non­range­compressed file format is that the
range compression helps to make one byte integers suf-
ficient to represent the data in image compression such
as JPEG file formats.
However, a number of new developments have:
made it possible to capture "undigital" pixel data, i.e.
as an array of REAL *8 (64 bit);
made it practical to process the data (the proliferation
of 64 bit computer architectures);
made such capture desirable, and such data useful, as
for example, in imaging where it is desired to combine
multiple exposures of the same subject matter. In this
case, for applications where linearity (homogeneity and
superposition) is desired, it is preferable that the image
data also be linear in the quantity of light.
Typically, because of the arithmetic involved in such
processing, it is desired to have REAL *8 (double pre-
cision, i.e. 8 byte floats) rather than merely REAL *4
(single precision).
Thus, in raw form, there is an eightfold increase from
INTEGER *1 (single byte integers) to REAL *8. Thus
there is an even greater need for image compression
when capturing "undigital" pictures.
2. 3 "Undigital" pictures that are also photo-
quantimetrically linear
We propose to store images with no range compres-
sion whatsoever. Ideally, the values encoded in the file
are simply measurements of the quantity (in the quan-
timetric sense [3]) of light present at each pixel element
of the sensor over a particular period of time.
This description implies representation as greyscale
images, but the idea is easily expandable to red, green
and blue sensitivities or other multibanded, multispec-
tral and color images.
In terms of redisplaying the
archived images, the display device simply needs to out-
put the red, green and blue intensities collected by the
sensor and archived in the file. There is no calibration
needed (once the output of the device has been cali-
brated) and consequently an accurate representation of
light is portrayed by the display device. The dynamic
range compressor and expander cannot affect the image
because they have been removed from the process.
However, the main benefit, beyond merely a true and
accurate display, is that processing done on the quan-
timetric data is really linear processing. Thus, for ex-
ample, deblurring done on the data so-represented is
deblurring in lightspace, rather than homomorphic fil-
tering. (It has been shown [2] that so-called linear filters,
when used on images, are not at all linear, and are in
fact incorrect homomorphic filters.)
background image
2. 4 Processing of traditional image formats
When video signals are processed using linear filters,
there is an implicit homomorphic filtering operation on
the photoquantity (a measure of the quantity of light
present on a sensor array element [3]). As should be
evident from Fig. 1, operations of storage, transmission,
and image processing take place between approximately
reciprocal nonlinear functions of dynamic range com-
pression and dynamic range expansion.
Many users of image processing methodology are un-
aware of this fact, because there is a common miscon-
ception that cameras produce a linear output, and that
displays respond linearly. Also, it is perceived that the
nonlinearities in cameras and displays arise from defects
and poor quality circuits, when in actual fact these non-
linearities are fortuitously present in display media and
deliberately present in most cameras. Thus the effect of
processing signals such as f
in Fig. 1 with linear filter-
ing is, whether one is aware of it or not, homomorphic
filtering. Tom Stockham advocated a kind of homomor-
phic filtering operation in which the logarithm of the
input image was taken, followed by linear filtering (e.g.
linear space invariant filters), followed by taking the an-
tilogarithm [5].
In essence, what Stockham did not appear to realize,
is that such homomorphic filtering is already manifest
in simply doing ordinary linear filtering on ordinary pic-
ture signals (whether from video, film, or otherwise). In
particular, the compressor gives an image f
= f (q) =
= q
(ignoring noise n
and n
) which has
the approximate effect of f
= f (q) = log(q + 1) (i.e.
roughly the same shape of curve, and roughly the same
effect, example: to brighten the mid­tones of the image
prior to processing). Similarly, a typical video display
has the effect of undoing (approximately) this compres-
sion, e.g. darkening the mid­tones of the image after
processing with ^
q = ~
) = f
. Thus in some sense
what Stockham did, without really realizing it, was to
apply dynamic range compression to already range com-
pressed images, then do linear filtering, then apply dy-
namic range expansion to images being fed to already
expansive display media.
2. 5 Correcting the nonlinear camera response
There exist certain kinds of image processing for
which it is preferable to operate linearly on the pho-
toquantity q. Such operations include sharpening of an
image to undo the effect of the point spread function
(PSF) blur of a lens, or to increase the camera's gain
retroactively. We may also add two or more differently
illuminated images of the same subject matter if the
processing is done in photoquantities. What is needed
in these forms of photoquantigraphic image processing
is an anti­homomorphic filter. The manner in which
an anti­homomorphic filter is inserted into the image
processing path is shown in Fig. 2.
Previous work has dealt with the insertion of an anti­
Figure 2. The anti­homomorphic filter : Two new
elements ^
and ^
f have been inserted, as com-
pared to Fig. 1. These are estimates of the the
inverse and forward nonlinear response function of
the camera. Estimates are required because the
exact nonlinear response of a camera is generally
not part of the camera specifications. (Many cam-
era vendors do not even disclose this information
if asked.) Because of noise in the signal f
, and
also because of noise in the estimate of the cam-
era nonlinearity f , what we have at the output of
is not q, but, rather, an estimate, ~
q. This sig-
nal is processed using linear filtering, and then the
processed result is passed through the estimated
camera response function, ^
f , which returns it to a
compressed tone scale suitable for viewing on a typ-
ical television, computer, or the like, or for further
homomorphic filter in the image processing chain. How-
ever, in the case of using a camera in which the raw
12-bit data is available, processing using the raw data
(NEF files), may proceed as shown in Fig. 3.
Figure 3. A modified method of photoquantimetric im-
age processing (shown in figure 2), in which the
raw data is available, and consequently no anti­
homomorphic filter is necessary. Moreover, a com-
parison (e.g.
comparametric analysis) between
compressed and raw data is possible.
3. Applications to file formats
As mentioned in the previous section, an unfortunate
consequence of using dynamic range compression is that
often the compressors and expanders don't match. Even
if they do match, there is still the problem that working
with images typically makes linear filtering impossible,
since we usually do not know the proprietary camera
response function that a camera vendor has used.
Typically, camera sensors are laid out in a Bayer pat-
tern[1] using red, green, and blue colour filters in order
to produce colour images. The sensitivity to red, green,
and blue wavelengths varies in different cameras. This
leads to differing pixel values in the dynamically com-
background image
pressed image. Furthermore, displays (monitors, pro-
jectors, and printed matter) vary in terms of their rep-
resentation of red, green and blue. These differences
lead to inaccuracies in representation, in addition to the
aforementioned inability to do true and accurate lin-
ear filtering. Often, this is noticeable in digital televi-
sions playing digital video. For example, in playback
of DVD recordings the red channel is often incorrectly
displayed, and obviously out of balance with the green
and blue channels. This contradicts the intuition that
digital media is necessarily better.
3. 1 Being "undigital"
The difference between discrete time signal process-
ing and digital signal processing is whether or not
the elements in a discrete lattice of samples are quan-
tized. Thus, "digital" implies quantization (finite word
length), and thus quantization noise, as well as the afore-
mentioned artifacts of such noise having a homomorphic
effect. Additionally, spatial resolutions of modern cam-
eras have increased dramatically, and will likely continue
to increase exponentially. As a result, spatial sampling
will soon surpass (and has in some cases already sur-
passed) the Nyquist rate, dictacted by the rest of the
system, such as lens, optics, etc., so that spatially, the
image is sufficiently "undigital" that the tonal quanti-
zation becomes the main limiting factor.
Accordingly, what is desired is an "undigital" file for-
mat that can better represent analog (continuous real-
valued) values, to which a REAL *8 representation pro-
vides a sufficiently good approximation, i.e. sufficiently
"undigital" as to, when combined with linearity, per-
mit true and accurate linear image processing. With
the advent of 64 bit (8 byte) computer architectures,
"undigital image processing" is now practical.
If the response of a camera is known, or the im-
age is recorded in lightspace as photoquantities, know-
ing the response of the display device permits images
to be displayed accurately.
With computers moving
to 64-bit processors (like the Apple Mac G5, AMD's
Athlon 64 and Itanium) double precision values may be
dealt with natively allowing photoquantities to be easily
handled by computers. Rather than using pixel based
storage such as portable pixmaps (PPMs) or jpeg
ages or PNGs, we have proposed photoquantimetric file
formats, specifically PLMs (Portable Lightspace Maps)
and JLMs (Jpeg Lightspace Maps). Such file formats
will allow for the efficient storage of accurately repre-
sentable images, as well as the capability for true linear
3. 2 The PLM file format
We began by extending the Portable aNyMap (PNM)
file format from its original PGM (portable grey map,
INTEGER *1 greyscale) and PPM (portable
pixmap, i.e. INTEGER *1 color), to include a REAL
*8 image format, as well as a REAL *8 linearly quanti-
metric format.
Thus we have added four new types, P7, P8, P9, and
PA to the existing P1-P6 types already defined:
P1 and P4
ASCII text and binary bitmap images
P2 and P5
ASCII text and binary greyscale images
P3 and P6
ASCII text and binary rgb colour images
P7 and PA
ASCII text and binary dynamically de-
compressed greyscale images (new)
P8 and PB
ASCII text and binary dynamically de-
compressed rgb colour images (new)
We also then developed a 64-bit file compression for-
mat similar to jpeg for dynamically decompressed high
dynamic range images of type P7 to PA.
3. 3 The need for compression of "undigital" im-
Typically, images are very large when stored in our
PLM file formats (i.e. 8 times larger). For example, a
640x480 colour PLM is approximately 7 megabytes. In
contrast, most images today use just 24-bits per pixel
which accounts for the standard 16 million colors (255
In this sense, PLMs are represented in 192-bit color. In
its raw form, the 192-bit color image is, in many situa-
tions, not practical, but when compressed, will often be
no larger than an ordinary JPEG image, because much
of the oversize is informatically redundant. Therefore,
the JLM (Jpeg Lightspace Map) could be used in many
forms of media so that a simple and intuitive standard
could be used for all imaging purposes.
4. Overview of algorithm
As mentioned, care was taken to follow the widely
accepted and used JPEG algorithm. This was done be-
cause of the algorithm's simplicity and proven reliability.
The major difference between a JLM and JPEG is in the
use of double precision numbers for JLM images. The
open source library set out by the joint photographic
experts group is based on INTEGER *1, and hence is
heavily optimized towards that one byte size. Whereas a
standard PLM file allows each pixel to carry a quantity
of light (q) that can range from 0 to the maximum al-
lowable double value. This flexibility easily handles any
desired precision required today and that of any image
in the foreseeable future. The main steps in the com-
pression process are outlined in Fig. 4. As described, the
compression is based on the use of quantizing (or elim-
ination of less important data) and that of run length
encoding (a simplified version of Huffman coding).
4. 1 Discussion of key parts
The construction of quantization tables
are key to the success of any jpeg-type compression rou-
tine. In the case of the JPEG Lightspace Map a base
equation utilizing a dynamic quality factor was used to
background image
8x8 Blocks
RLE tables
Figure 4. Schematic of JPEG compression: A sim-
ple block diagram depicting the process. The image
is first broken up into 8x8 blocks as to minimize
the strain on the DCT process. From there each
block is DCT encoded. After applying quantization
to these matrices, run length encoding following a
zigzag pattern is written to a table in the final file.
The summation of all the 8x8 blocks makes up the
final JLM file.
determine the exact quantization values. Higher values
in a table result in good compression and poor qual-
ity and lower values result in less compression in higher
quality. A compromise between the two was found as to
be discussed in the experiment section.
Run Length Encoding
If the quantization tables used
in the quantization process worked well, the resulting
8x8 blocks should be plentiful with zeros.
The run
length encoding that follows a zigzag pattern (to capital-
ize on the tendencies of the DCT), uses simple encoding
logic to prevent the storage of less important informa-
tion that has been reduced to zero from being stored
into 8-bytes. The exact structure and developed stan-
dard will be discussed later in this paper. The "throwing
out" of information is simply the removal of the elements
of the DCT with a low-bin count and/or high frequency.
This can have adverse effects in images with many sharp
lines and text, as what is presumed to be noise in this
instance could be texture.
5. Results
An example of the compression is shown in Fig. 5. It
depicts a typical lightspace image and how the different
quality factors affect the appearance of the image.
The graph shown in Fig. 6 gives us a sense of the
amount of compression we obtain when working with a
typical lightspace image. It was found after experiment-
ing with a variety images that a quality factor of (7) on
a scale from (1-12) produced a compressed image with
comparable quality to the original.
When using sharp lined images and text the JLM
compression routine tended to blur these edges and re-
duce quality notably. It is obvious that user discretion
is needed to acknowledge a higher quality factor should
be used on sharp images, while more scenic and flowing
images can do with a smaller factor.
5. 1 JLM file syntax
It was attempted to keep the syntax as simple as
possible so manipulation and reading is quick and easy.
Header :
Figure 5. Varying levels of compression :
are six different levels of compression on the same
image. The original image (a) is compressed to
form the other five images. It should be quite clear
the increasing compression (or decreasing quality
factor) moves left to right and downwards across
the figure. We can start to see a notable difference
in the images at (d) which is at a quality factor
of (5). As the compression increases we can see
that a blocky effect appears. Also, early on in the
compression process detailed sections of the eye-
tap worn by Professor Mann (the one in the fore-
ground) show blurring. This is an example of the
utility having difficulty with sharp lines and sud-
den color change. However, if you look at Professor
Mann's forehead, where there is little color change,
we can see little difference between the original and
further compressed images.
image identifier
JLM for JPEG Lightspace Map
height and width
Picture dimensions
quantization tables
Equation used to obtain (involv-
ing array indicies)
quality factor used
Huffman table (for each 8x8 block) :
number of non-zero numbers in matrix
number of zeros preceding non-zero value
non-zero value
6. "Undigital photography" as a form of
visual art
The proposed image representations are also useful
for the production of visual art, such as simulation of
multiple exposures, or accurate simulation of other pho-
tographic phenomena.
For example, the proposed image representations
were used to create various multiple exposure art forms,
background image
Percentage size of original PLM
Quality factor (1-12)
Compression ratio of JLM
Figure 6. Percentage Compressed vs. Quality Fac-
tor :
We can see that after a certain threshold
(ten in this instance) the rate of compression is
small. To capitalize on this one should not apply a
smaller quality factor as the compression benefits
will be quite small in comparison.
and cement these together into a single image. Such
combining differently illuminated exposures of the same
subject matter, is illustrated in Fig. 7.
7. Conclusion and future work
In this paper, we have proposed two file formats for
standardized storage, transmission, and processing of
quantimetrically linear image data, using REAL rather
than INTEGER quantities.
The first of these, the
Portable Lightspace Map (PLM), is an extension of
the PNM file format, but the file sizes are quite large.
The second, the Jpeg Lightspace Map (JLM) is a com-
pressed version of the PLM that uses a data compres-
sion methodology that builds upon JPEG compression.
Both of the two proposed file formats are intended to
be used independently, without dynamic range compres-
sion or expansion, and are therefore truly universal im-
age formats that permit quantimetrically linear image
processing. The formats are independent of the partic-
ular camera that takes the picture and the particular
device that displays them, and hence allow the user to
reclaim the original photoquantities with ease. Also, as
the formats are based in lightspace they provide an in-
tuitive and easy to use foundation for intelligent image
processing. With the creation of the JLM, lightspace
file sizes are manageable for the everyday user and as a
result these formats can now more easily be adopted by
the computing community. Current work on the project
is focused on optimizing and integrating the JLM com-
pression utility to perform on 64-bit based processors
and run symbiotically with other lightspace utilities al-
ready in use.
[1] B. Bayer. Color imaging array, 1976. U.S. Patent
[2] S. Mann.
Comparametric equations with practi-
Figure 7. Various exposures to different sources of illu-
mination are combined quantimetrically. In, for ex-
ample, v32.jpg, the open basement door under cell
block "A" on Alcatraz Island, is exposed to light
from a flash lamp held to the left. The flash lamp
is then moved to the right, to illuminate the scene
from the right, in exposure v36.jpg. Finally, expo-
sure v37.jpg captures light coming from upstairs,
beyond the jail bars above the door. Each expo-
sure is made by illuminating the space from one of
various viewpoints. These pictures are then con-
verted into lightspace by applying an estimate of
the inverse of the camera's photographic response
function[2]. The resulting photographic quantities
are added together, and the combined exposure is
then converted from light space back into a picture.
cal applications in quantigraphic image processing.
IEEE Trans. Image Proc.
, 9(8):1389­1406, August
2000. ISSN 1057-7149.
[3] S. Mann. Intelligent Image Processing. John Wiley
and Sons, November 2 2001. ISBN: 0-471-40637-6.
[4] C. Poynton. A technical introduction to digital video.
John Wiley & Sons, 1996.
[5] T. G. Stockham, Jr. Image processing in the context
of a visual model. Proc. IEEE, 60(7):828­842, July