Object Recognition in Compressed Imagery.

Citation

Daniell, Cynthia Evors (2000) Object Recognition in Compressed Imagery. Dissertation (Ph.D.), California Institute of Technology. doi:10.7907/6z3z-ar86. https://resolver.caltech.edu/CaltechTHESIS:10052010-135012445

Abstract

It is often necessary to search for objects in large databases of compressed imagery. In the past, object recognition and image compression have generally been treated as separate problems, resulting in inefficient suboptimal performance. Moreover, computational and storage issues make it fundamentally prohibitive to uncompress large images prior to object recognition. We provide two complementary solutions to the problem of object recognition in compressed imagery, each of which integrates subband and correlation filtering in a unique manner.

One key benefit of correlation filters is that, as linear systems, they are highly compatible with the subband filtering process. This enables us to provide a seamless operation in which object recognition and data compression are viewed as continuations of the same process. The public MSTAR data set illustrates our results on a three class problem of 79 Synthetic Aperture Radar images at one foot resolution.

Our general framework, the Pattern Recognition Subband Coder (PRSC), provides simultaneous synthesis and recognition at full resolution in a computationally efficient architecture. Its parallelism enables a result 1.6 times faster, in the limit, than correlation on uncompressed imagery. Furthermore, by jointly optimizing the synthesis and recognition filters, the PRSC achieves 100% recognition accuracy on our compressed data set, improving performance over that produced from the original (uncompressed) data set, by 3.7%. We maintain this success for compression ratios up to 6:1.

Addressing the issue of reduced resolution recognition, our Subband Domain Correlation Filters operate directly on the subband coefficients at multiple resolution levels. For compression ratios of at least 20:1, we achieve recognition performance of at least 90%, 85%, and 75%, respectively, on two, four, and eight foot resolution data.

Thus, through our solutions with compressed imagery, we outperform correlation results on the equivalent original imagery in terms of both speed and accuracy, as well as provide success at reduced resolutions of the data.

Thesis Files