International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B3. Istanbul 2004
Figure 1 is a flow diagram of JPEG compression and
decompression.
(1) Color space conversion
A source image represented in the RGB color space is
converted into an image represented in the YCbCr color
space in JPEG compression.
Inversely, the transformation from the YCbCr color space
to the RGB color space is executed in JPEG
decompression.
(2) Downsampling / Upsampling
Data reduction of chrominance components Cb and Cr is
performed in JPEG compression as required. Most of
pieces of JPEG compression software adopt the 4:1:1
sampling as the default setting. The 4:1:1 sampling means
sampling four data of the luminance component Y and one
data for each chrominance component Cb and Cr from two
horizontal pixels by two vertical pixels. The 4:4:4
sampling is also used, which means no data reduction of
chrominance components.
Restoration of chrominance components Cb and Cr to full
size is performed in JPEG decompression if necessary.
(3) Forward DCT / Inverse DCT
The source image is transformed into the spatial frequency
domain by DCT in JPEG compression.
The compressed image data in the spatial frequency
domain is transformed into the image domain by inverse
DCT in JPEG decompression.
(4) Quantization / Dequantization
DCT-coefficient data are quantized in JPEG compression.
The quantization table used in this step controls the
compression level of an image. The quantization stage is a
lossy process.
Dequantization in JPEG decompression is performed by
using the quantization table included in the compressed
image data.
(5) Entropy encoding / Entropy decoding
Arithmetic entropy encoding is performed to further
reduce compressed image data volume in JPEG
compression. Huffman coding is a commonly used
method.
Arithmetic entropy decoding is performed in JPEG
decompression.
Figure 2. Test image (digitized color aerial photograph)
648
In JPEG compression, the compression level of an image can be
controlled by a constant, which is generally called the quality
factor (Q-factor). Q-factor sets the quantization table. A higher
Q-factor gives higher compression. A lower Q-factor gives a
better quality image, but a lower compression ratio. Therefore,
variable compression can be achieved by simply scaling the Q-
factor. An important property of the JPEG compression scheme
is the adjustment of the Q-factor to balance the reducing image
size and degraded image quality. In fact, different JPEG
compression programs have different Q-factors
2.2 Evaluation of Image Quality of Reconstructed Images
RMSE and PSNR are usually used to compare various image
compression techniques (Santa-Cruz et al., 2000, Grgic ef al.,
2001, Taubman et al., 2002). Rountree et al. (2002, 2003)
demonstrated the effectiveness of JPEG 2000 compression for
remote sensing data by PSNR.
Fukue et al. (1998, 2000) evaluated the effects of lossy
compression of a pair of stereo satellite images by the accuracy
of surface measurement. Moreover Li ef al. (2002) investigated
the effects of lossy compression on the accuracy of
photogrammetric point determination.
Numerical evaluation of quality of a reconstructed image after
decompression is seldom conducted. Al-Otum (2003) reported
the experiment on various numerical measures to determine a
proper measure of image quality of a reconstructed image.
Since most of studies on assessment of lossy image
compression utilized monochrome images and no texture
measures for evaluation, results of the studies cannot be applied
directly for GIS application. Therefore, we decided to carry out
an empirical investigation into the effects of image compression
on quality of color aerial images by using color and texture
measures.
3. EXPERIMENT
3.1 Test Image
A color aerial photograph of a city taken at a scale of 1:7000
was digitized. Scanning resolution was 40 um on the
photograph, that is, 280 mm on the ground. Each color
component (Red, Green and Blue) was quantized into 8 bits
(256 levels). The test image of 4096 pixels by 4096 lines was
extracted from the center part of the scanned image. Figure 2
Image size 4096 pixels x 4096 lines
Gray level features [R] [G] [B]
e mean 149.2 141.3 142.7
e standard deviation 75.1 63.4 48.0
Color features [H] [S] [V]
e mean 110.6 55.5 162.3
e standard deviation 75.9 27.2 63.9
GLDV [R] [G] [BI =
e contrast 18.8 18.4 15.7
e angular second moment 0.220 0.158 0.154
e entropy 3.92 4.13 4.06
e mean 10.8 11.7 10.7
Fourier power spectrum [R] [G] [B]
e mean spatial frequency 110.1 132.0 153.6
Table 1. Statistics of test image
Int
VI
show
test i
We c
was |
inves
locat
comy
volur
other
was t
and |
Bloc]
statis
of Bl
321
Four
Three
comp
2000