Segmentation
The picture is first divided into segments of
equal sizes. The number of horizontal divisions is kept equal to
the number of vertical divisions, so that the size of the image is kept
equal. Each segment is then divided into 6X6 windows. Since
our images are 480X640, we divided our image into 4X4 segments, removing
part of the image to give an overall image that is 480X624.
Statistical Measures for each window
For each window, the following statistical measures
are calculated:
i) mean pixel intensity M(W)
L = number of rows
M = number of columns
W(l, m) = value of pixel at the
lth row and mth column
ii) mean absolute deviation D(W)
iii) horizontal edge count ECh(W)
THLD(x, y) = 1 (if x>y) or 0
(if x<=y)
T(W) = threshold proportional to
the deviation = lambda*D(W)
lambda = 0.75 (range from 0.8 to
1.2)
To qualify as an edge, the change
in pixel intensities must exceed this threshold.
iv) vertical edge count ECv(W)
v) horizontal mean edge magnitude EMh(W)
CLIP(x) = x (if x>0) or 0 (if
x <=0)
vi) vertical mean edge magnitude EMv(W)
vii) normalized mean edge magnitude EM(W)
Measures for each segment
After the measurements are taken for each window,
the mean and mean absolute deviation of each measure is calculated over
the entire segment.
Features and Visual Significance
Focus, contrast and texture are used in evaluating the visual significance
of a segment.
Contrast refers to the range of tones present in an image. High contrast
regions tend to have a strong visual impact. This is because a high global
contrast means better isolation of the subject from the background.
Also, high local contrast reveals details more clearly in sharp areas.
A simple way to measure contrast is to calculate the deviation of pixel
intensities in a region.
Focus represents the intent of the photographer, and is hence an important
visual feature. Regions in focus are likely to have crisp or sharp edges
characterized by steep transitions, and vice versa. Even though they are
strongly related, contrast and focus are distinct features. Sharp regions
tend to have a higher contrast than blurred regions. However there are
situations when a region can have a high contrast even if it is out of
focus. A simple way to measure focus of a segment is to find the average
normalized mean edge magnitude over the entire segment.
Measurement of sharpness takes into account of the contrast charateristics
of the segment, since contrast is measured in terms of deviation, which
in turn is used in qualifying a transition as an edge.
Texture refers to the surface nature of objects. It is seen on an image
as a repetitive variation of tone and detail. To detect the presence or
absence of textures, the texture of a segment S is as follows:
where the first term is the reciprocal of the segmental deviation
of the deviations in the windows, the second term is the reciprocal of
the minimum of the segmental deviation of the horizontal and vertical edge
counts in the windows. In a textured region with repetitive details, these
segmental deviations can be expected to be small, resulting in a high value
for T(S).
After C(S), F(S) and T(S) are computed for an image, they are normalized
so that their range is from 0 to 1. A linear combination of the contrast
and texture values is multiplied by the focus to obtain visual significance,
V(S), of a segment S as shown below:
where a is a real number between 0 and 1 with typical values ranging
from 0.50 to 0.75. We chose a value of 0.7.
To identify the visually significant segments within an image, a visual threshold VT, 0.2, is chosen. Segments with V(S)> VT are considered to form the subject. Segments with V(S) < VT are considered to form the background.