Improved Grouping and Noise Cancellation for Automatic Lossy Compression of AVIRIS Images - Advanced Concepts for Intelligent Vision Systems - page 263

PSNRnf PSNRor

PSNR

46

44

42

40

38

36

34

32

30

28

QS

4 8 12 16 20 24 28 32 36 40 44 48

Fig. 2. PSNR or ( QS )and PSNR nf ( QS ) for the one-channel Frisco test image

zation step is to be set as QS =4 . 5 σ . Since the estimate σ is not absolutely
accurate, some oversmoothing or undersmoothing of the noise due to the lossy
compression might take place. The undersmoothing that is observed for ˆ σ<σ
is less dangerous, but then the attained compression ratio is smaller than can
be reached for QS =4 . 5 σ .
The existence of the image-dependent noise in AVIRIS-images and the noise-
dependent lossy compression is demonstrated in Fig. 3. In Fig. 3, a) there is a
sample of a noisy channel from the image Cuprite, band 2. Then the band is
compressed/decompressed with the quantization step QS ( n )=4 . 5 σ n where n is
the index for a sub-band. The reconstructed band is in Fig. 3, b). The denoising
is now well observed. In Fig. 3, c) there is the band 30 from the Cuprite image
having high SNR. In turn, Fig. 3, d) demonstrates the decompressed image. As
seen, no distortions (losses) are observed visually. Thus, we can state that for the
sub-bands with rather low SNR the proposed approach to the lossy compression
provides ecient denoising whilst the useful information contained in the sub-
band images characterized by a high SNR is preserved well.
AVIRIS images have different contents depending on the area under imaging.
In Fig. 4 two bands from the Moffett Field image are shown. On left, there are
the original images and on right, there are the compressed/reconstructed images.
On top row, the band 128 is shown, on bottom row, the band 223 is shown. Com-
pared to the Cuprite image in Fig. 3, there are more tiny details in the Moffett
Field image. From the compression point of view, the more complex structures
mean a lower compression ratio. From the denoising point of view, the proposed
approach is operating similarly to a less detailed image: the quantization step is
still determined from the noise characteristics only.