[2 004 International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B7. Istanbul 2004
one Tt.
Ref = i
» the ; | ;
} FIDELITY = of (HH
18,8, 18g,
hart (2) Compression to the reference image
2
At first, the 3D wavelet transformation is done to the reference
IN image. The lossless compression is adopt to the low frequency PSNR = 101g a = = 48 — 201g6 (12)
wavelet coefficient. And the arithmetic coding is taken to
compress the high frequency wavelet coefficient.
(3) Compression to the i spectrum image where c . isthe pixel grey value of original image.
J
Decompose the i spectrum image using the 3D wavelet
transformation. The lossless huffman coding is adopted to the g _ is the pixel grey value of the reconstruct image.
Ge low frequency image part. Then edge is extracted from the high 7
wavelet coefficient image and be compression using Huffman MN is the pixel number of image
^ 1 a 2 ^ edoe 1 © a 1 a > . . .
coding. Remove the edge image from the high frequency 3 is the standard deviation and.can be calculated. as
wavelet coefficient image, the left wavelet coefficient is
quantification with the specific level. In this paper, 4 level is
taken to quantized the left wavelet coefficient. Finally, the
following equation.
d ven = (13)
arithmetic coding is done to the quantification wavelet =
coefficient image.
(4) Repeat (3) till all the multispectrum images are compressed. Fidelity is the geometry distortion of the reconstruction
The corresponding compression images reconstruct flow chat is image comparing to the original image. PSNR (peak
illustrated as the figures : signal noise ratio) is the radiation distortion of the
| | reconstruction comparing to the original image.
ode] Compressed Compressed CORE CO Nd Obviously, if the image has no loss, the fidelity is 1.0 and
ate ta ate . . . > :
Data 2.of Data Loi aco Bt o PSNR is infinity (expressed as & =0, & is the gray
Reference Band| [Reference Band Band i Band i MS Y
Y i i i standard deviation of the homologous pixels between the
ed Tr ATI AEN o original image and the decode image ). If 0 equals 1.0,
^5se : rithmetic rithmetic uffman ; à 8 4 ; :
M Decode Prods Decade Decods PSNR is 48. If 6 equals 2.0, PSNR is 42. The quality
li v Y Y + request of the quassi-lossless compression comes from
E re = : v the standard deviation above 2.0, which ensure not to
Band Reference Band Reference Quantification || Band i Low xS uis
Low Frequency High Frequency Restoration Frequency affect the application.
Coefficient Coefficient Transformation | | Coefficient
| J Table 3 and Table 4 show the compression result of the
—À v Edge Extraction || Band i Keep High test images using the proposed method in this paper with
3D Inverse Wavelet frequency coefficient reference image.
Transformation without Edge ; ; ; ;
Table 3 Compression Result Using Low Correlation Reference Image
Y Edge Image of
Images Reference Band [—— Total Reconstruction
Reference Image Reference Average Compression Images
Reconstruction Band i High Name Band Correlation : ah B
sion Image Frquency atio Quality
Coefficient TM 2 0.626 12.32 0.9901/41.36
| E 2 0.735 13.15 0.9897/41.86
Y MODIS 11 0.32 11.37 0.9756/40.91
y the 3D Inverse Wavelet
ratio Transformation Table 4 Compression Result Using High Correlation Reference
the Total
lyse. + : Image Reference Average COR io Reconstruction
ition Reconsihuction Name Band Correlation ati Images Quality
nage mane TM 3 0.941 17.42 0.9992/46. 1 1
ined E 16 0.946 16.78 0.9984/44.52
Aon [tnd 1 MODIS 3 0.981 18.54 0.9987/42 87
Figure 5 Reconstrunction Technique Flow
As table 3 and table 4 shown, the multispectrum images
3. EXPERIMENTS AND RESULTS compression quality are based on the reference image
selection. The maximize correlation spectrum images
Fidility and peak signal noise ratio(PSNR) are taken as the have the biggest redundancy . Thus, compression can
quality assessment parameters of the reconstruct image. obtain high compression together with high
reconstruction image quality.
59
