Page 254 - Digital Analysis of Remotely Sensed Imagery

216 Cha pte r S i x

between DN 2 and DN 3, between DN 8 and DN 9, between
DN 10 and 11, and between DN 14 and DN 15.
• Second, the disparity between the maximum frequency
(9 pixels) and the minimum frequency (1 pixel) has been
reduced. Now the frequency ranges from 2 to 9 pixels.
• Finally, the spectral distance between any two pixel-containing
DNs has been enlarged. In the input image, this distance is
invariably 1. However, it varies from 1 to 3 in the histogram-
equalized image. Moreover, the higher the frequency at a DN,
the broader the distance. Therefore, the most dominant pixels in
the input image are enhanced more than minority pixels. They
are more easily perceived in the histogram-equalized image.
Identical to truncated linear stretching, histogram equalization
always involves loss of information. However, the manner of loss dif-
fers widely. Unlike truncated linear stretching in which the loss is
always restricted to the darkest and brightest pixels, in histogram
equalization the loss can occur at any DN level if there are few pixels
at this level. In other words, the loss takes place at whatever gray
level so long at it has a minority of pixels. Furthermore, the degree of
stretching is proportional to the frequency. Those DN levels with a
higher frequency are stretched more than those with a lower fre-
quency. This explains why the interval between any two vertical bars
in Fig. 6.8b is not uniform.
Histogram equalization is an effective means of contrast
enhancement. A high degree of stretching is achieved at the expense
of losing information for minor pixels. As illustrated in Fig. 6.9, both
the raw and enhanced images have the same DN range from the
minimum of 0 to the maximum of 255. However, the raw image has
a low contrast with most of the pixels having a value confined to a
narrow range of DNs (Fig. 6.9a). They lean toward the lower end of
the DNs, causing the image to have a rather dark tone. In its histo-
gram-equalized counterpart, the dominant pixel values have shifted
to the midrange, and the DN range has been extended. Conse-
quently, more details are visible. Since the histogram is a bell-shaped
curve, the loss of information is restricted to the lowest and the
highest DNs. Namely, the darkest and the brightest features become
indistinctive in the output image.

6.2 Histogram Matching
The tonal inconsistency problem in creating a mosaic from multiple
aerial photographs (Sec. 2.7) can be eliminated or reduced to a lesser
extent through histogram matching. The principle underlying histogram
matching is rearrangement of the pixel values in a slave image in such
a way so as to achieve a distribution approximately identical to that of