Share Email Print
cover

Proceedings Paper

Correlation pattern recognition in compressed images
Author(s): Abhijit Mahalanobis; Cindy Daniell
Format Member Price Non-Member Price
PDF $14.40 $18.00

Paper Abstract

This talk deals with correlation filtering techniques developed for finding patterns in Wavelet compressed imagery. It has been shown that the correlation filters can recognize patterns in IR and SAR imagery at very high compression rates in excess of I 00 to 1. The reason is partly due to the excellent information compaction capability of wavelets augmented by the zero-tree encoding technique, and partly because correlation filters do not require pixel level reconstruction of visual information, but rather the preservation of spectrally significant information which wavelet encoding may achieve very well.

As part of the presentation, we will describe a technique for merging the inverse wavelet compression and correlation filtering operations into a seamless process. In addition to being theoretically elegant, this has the added benefit of reducing the overall computations. Equally significant, it offers a method for obtaining the correlation result without requiring the full image to be reconstructed thus avoiding the need for large amounts of storage. We also show that it is indeed possible to design the compression filters and the correlation filter in a joint optimization process with added benefits. The notion of performing recognition by directly exploiting the Wavelet coefficients is also addressed. Here, we describe a technique which combines the information in different bands using a multi-channel correlation algorithm known as polynomial correlation filters. The optimization process must take into account the shift-sensitivity of wavelet coefficients. It is shown that simultaneous optimization of the sub-band QMFs and the correlation filters leads to promising results.

Paper Details

Date Published: 30 November 2001
PDF: 22 pages
Proc. SPIE 10302, Optoelectronic Information Processing: Optics for Information Systems: A Critical Review, 1030208 (30 November 2001); doi: 10.1117/12.449687
Show Author Affiliations
Abhijit Mahalanobis, Lockheed Martin Corp. (United States)
Cindy Daniell, California Institute of Technology (United States)


Published in SPIE Proceedings Vol. 10302:
Optoelectronic Information Processing: Optics for Information Systems: A Critical Review

© SPIE. Terms of Use
Back to Top