xpected that enormous
will be Offered, for
nount of data.
'd to other situations.
ian behavior there, we
lel by replacing System
? different sensors such
replacing observation
need to make better the
1 defined in section 3
ly, automatic human
time tracking. For this
ompleted as explained
<. Furthermore, we aim
strian behavior using
iman tracking in high-
or Intelligent Vision
estrian choices, In H.
vior: Models, Data
erald Group, 2009.
Analysis, Cambridge
A. F. M. 1993. Novel
'esian state estimation,
'eedings F, Vol. 140,
Monte Carlo mixture
sting, Proceedings of
Science and Electrical
lan-view map based
ind tracking, Pattern
5.
and Cruz, 1., 2009.
"a pedestrian walking
esearch Part B:
ion and tracking of
yesian combination of
Journal of Computer
on Inverse Problem,
ety for the Promotion
sts B (22760401).
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Volume XXXIX-B3, 2012
XXII ISPRS Congress, 25 August — 01 September 2012, Melbourne, Australia
FROM IMAGE CONTOURS TO PIXELS
G. Scarmana
University of Southern Queensland, Australia
Faculty of Engineering and Surveying
gabriel.scarmana@usq.edu.au
Commission III/5
KEY WORDS: Image processing, Image compression, Image contours, Image reconstruction
ABSTRACT:
This paper relates to the reconstruction of digital images using their contour representations. The process involves determining the
pixel intensity value which would exist at the intersections of a regular grid using the nodes of randomly spaced contour locations.
The reconstruction of digital images from their contour maps may also be used as a tool for image compression. This reconstruction
process may provide for more accurate results and improved visual details than existing compressed versions of the same image,
while requiring similar memory space for storage and speed of transmission over digital links.
For the class of images investigated in this work, the contour approach to image reconstruction and compression requires contour
data to be filtered and eliminated from the reconstruction process. Statistical tests which validate the proposed process conclude this
paper.
1. INTRODUCTION
An important area of digital image processing is the
segmentation of an image into various regions to separate the
objects from the background. Image segmentation can be
categorised by three methods.
The first method is based upon a technique called image
thresholding (Russ, 2007), which uses predetermined grey
levels as decision criteria to separate an image into different
regions based on the grey levels of the pixels. The second
method uses the discontinuities between grey levels regions to
detect edges within an image (Baxes, 1994). Edges play a very
important role in the extraction of features (and their
dimensions) for object recognition and identification in image
analysis operations (Gonzalez and Wood, 2008). The final
method relates to detecting edges so as to separate an image into
several different regions by way of grouping pixels that have the
same grey levels.
Since edges play an important role in the recognition of objects
within an image, contour description methods have been
developed that completely describe an object based upon its
contour. The most common methods are (a) a method based
upon a coding scheme called chain codes, (b) the use of higher
order polynomials to fit a smooth curve to an object's contour,
and (c) the use of Fourier transform and its coefficients to
describe the coordinates of an object's contour (Russ, 2007).
Apart from image analysis operations, edges and contours have
been used for image editing (Elder and Goldberg, 2001) and
image compression applications. For instance, Elder and Zucker
(1998) proposed a novel image compression scheme based on
edge detection techniques where only edge and blur data rather
than the entire image is sent over a digital link. This
compression technique is especially appealing because edge
density is linear in image size, so larger images will have higher
compression ratios. The argument is that for natural images,
most scenes are relatively smooth in areas between edges, while
large discontinuities occur at edge locations. Thus, much of the
information between edges may be redundant and subject to
increased compression. The work of Elder and Zucker proposes
two necessary quantities: the intensity at edge locations in the
image and an estimate of the blur at those locations. Using these
quantities, it is possible to reconstruct an image perceptually
similar to the original.
On the other hand, Vasilyev (1999) proposed a method for
compressing astronomical images based on simultaneous edge
detection of the image content with subsequent converting of
the edges to a compact chained bit-flow. Vasilyev combined
this approach with other compression schemes, for example,
Huffman coding or arithmetic coding, thus providing for
lossless compression schemes comparable to present
compression protocols such as JPEG.
The proposed reconstruction process uses image contour
detection rather than edge detection. The definition of contour
used in this work relates to the term used in cartography and
surveying where a contour line joins points of equal elevation
(height) above a given reference level (Smith et al. 2009).
In this context, the pixel values stored in an image can be
considered as the values of some variable z where each pixel
can be assumed as an elevation value z at its x and y
coordinates, thus defining a 3D shape (Weeks, 1996). This
shape is often a complex 3D surface that can be represented by
scattered contours nodes where each node (or vertex) of a given
contour also corresponds to a position (x,y) having a constant
colour intensity or elevation z.
For the class of grey-scale images investigated in this work (i.e.
human faces, landscapes and relatively small aerial images) this
reference level and the contour intervals (or contour increments)
are selected based on the dynamic range and/or on the image
histograms. Image histograms provide a convenient, easy-to-
read graph of the concentration of pixels versus the pixel
brightness in an image. Using this graph it is possible to discern
