Document 7189658
Download
Report
Transcript Document 7189658
Low-Complexity Lossless
Compression of Hyperspectral
Imagery via Linear Prediction
Fei Nan
& Hani Saad
By:
Presented to:
Dr. Donald Adjeroh
Index
Hyperspectral Images, what are they?
Remote Sensors and Low-complexity
Image Compression
Linear Prediction (LP)
Spectral Oriented Least Squares (SLSQ)
LP Implementation
SLSQ Implementation
Experimental Results
Improvements
References
Hyperspectral Image Compression
2
Hyperspectral Images
High-definition electro-optic
images
Used in surveillance, geology,
environmental monitoring, and
meteorology
224 contiguous bands
3 or more consecutive
scenes
Hyperspectral Image Compression
3
Remote Sensors & Low-complexity
Image Compression
Hyperspectral sensors measure
hundreds of wavelengths
Airborne vs. Satellite Sensors
Why low-complexity compression?
Hyperspectral Image Compression
4
Linear Prediction (LP)
Spatial correlation
Spectral correlation
LP
• Interband linear prediction for interband
coding
• Standard median predicton for
intraband coding
Hyperspectral Image Compression
5
Linear Prediction cont’d
Standard median predicton
• Used for intraband coding
Xi-1,j-1,k
Xi,j-1,k
Xi-1,j,k
Xi,j,k
Hyperspectral Image Compression
6
Linear Prediction cont’d
Interband linear prediction
• Used for interband coding
Hyperspectral Image Compression
7
Spectral Oriented Least Squares
(SLSQ)
Prediction defined in two different
enumerations for pixel:
1. Intraband enumeration
2. Interband enumeration
Hyperspectral Image Compression
8
LP Implementation
The first 2 conds apply to Interband. 2nd
cond can be skip when T=œ, given T gives
best performance.
The 3rd cond applies to Intraband(IB).
Hyperspectral Image Compression
9
SLSQ Implementation
The distance of Interband
and intraband are
defined.
The Predictor Error
Matrix C and Matrix X
The simplified form when
we assigned M=4 and
N=1.
Hyperspectral Image Compression
10
Experimental Results
Hyperspectral Image Compression
11
Experimental Results cont’d
128x128x224
LP
SLSQ
Cuprite
1.918 2.425
Jasper
1.850 2.364
Low
Altitude
1.708 2.131
Lunar
Lake
2.065 2.390
Hyperspectral Image Compression
12
Improvements
Using M=5 vs.
M=4
Keeping N=1
Future
improvements can
include look-ahead
prediction
SLSQ2 SLSQ1
Cuprite
2.443
2.425
Jasper
2.358
2.364
Low
Altitude
2.129
2.131
Lunar
Lake
2.413
2.390
Average
2.33
2.32
Hyperspectral Image Compression
13
References
Randall B. Smith, Ph.D., 17 September 2001.
MicroImages, Inc. Introduction to Hyperspectral
Imaging with TNTmips. www.microimages.com
Peg Shippert, Ph.D., Earth Science Applications
Specialist Research Systems, Inc. Introduction
to Hyperspectral Image Analysis.
Suresh Subramanian,, Nahum Gat, Alan Ratcliff
, Michael Eismann. Real-time Hyperspectral
Data Compression Using Principal Components
Transformation
Hyperspectral Image Compression
14