At the last 10 years the technology has grown, and gave us some instruments which were not available before.
Abstract
At the last 10 years the technology has grown, and gave us some instruments which were not available before. A good e.g. is the Hyper-spectral Imaging. By multi-spectral filming an object (e.g. brain, surface etc.) we can analyze its components through the spectral reflex. This project deals with analyzing Hyper-spectral Imaging using Blind Source Separation (BSS). The main objective of this project is to improve the results of an earlier project by using sparse representations of the data. The sparse representations gives us a better (simpler and still accurate) database to analyze the filmed surface without loosing important data.
The problem
Hyper-spectral Imaging gives us lots of information about the surface which was filmed. This data is normally under the effect of many factors as the camera resolution, noise made by the weather etc. Trying to analyzing this data without preprocessing can be done, but usually won’t give good results. Each pixel in a Hyper-Spectral cube represents about 100 by 100 feet surface, which contains several and different minerals. In order to separate the result to known minerals in a better way, we represented each mineral in a sparse and different (from the other minerals) representations. This representation, as shown in the project, allows us to separate the minerals much more efficiently.
The solution
Figure 1 – Block diagram of the sparse process.
The sparse representations was received from 9 different transforms, so each and every mineral that was represented by non-sparse 256 coefficients, had now 9472 sparse coefficients. At the next stage, we sorted the coefficients to select the ones that help us to increase the differences between the minerals. The final result allowed us to increase this differences 30 times better (!) than the original status using only 38 coefficients.
Our criterion for the separation between two minerals was:
Using BSS to separate the data with the sparse representations that we created gave, as expected, much better results:
Tools
Matlab 6.5 with two special toolboxes: Atomizer and Wavelab.
Those toolboxes are free to download from:
http://www-stat.stanford.edu/~wavelab/
http://www-stat.stanford.edu/~atomizer/
Conclusions
The main conclusion of this project is that the assumption that the results of the BSS can improve by sparse representations is correct. As it was mentioned, no test on a real hyper-spectral image was made, there for we recommend testing this method on a full and real database.
Other optional uses:
# Medical use – this method can be use on varied scan in order to find abnormal components as tumors.
# Military use – by scanning with Hyper-spectral camera from a plain or satellite a specified area, one can analyze its structure.
# Geographic & Financial use – analyzing the structure of a large scale surface can give us good knowledge whether it can be excavate.
Acknowledgment
First we would like to thank our project guide Mr. Ran Kaftory, for the time and patient over the last year. We are most certain that without his help and guidance we would not get to finish this project.
We would like to thank also to Mr. Johanan Erez, the chief engineer of the VISL lab for the administration work on this project.
Special thanks go to the Ollendorff Minerva Center that supported this project.