2026-09-19T11:19:03Zhttps://dspace.mit.edu/server/oai/requestoai:dspace.mit.edu:1721.1/748902022-01-13T07:54:29Zcom_1721.1_7582com_1721.1_7581col_1721.1_131022
William T. Freeman.
Chang, Hyun Sung
Massachusetts Institute of Technology. Dept. of Electrical Engineering and Computer Science.
Massachusetts Institute of Technology. Department of Electrical Engineering and Computer Science
2012-11-19T19:15:53Z
2012-11-19T19:15:53Z
2012
2012
http://hdl.handle.net/1721.1/74890
813987999
Thesis (Ph. D.)--Massachusetts Institute of Technology, Dept. of Electrical Engineering and Computer Science, 2012.
Cataloged from PDF version of thesis.
Includes bibliographical references (p. 145-156).
Compressed sensing is a recent theory for the sampling and reconstruction of sparse signals. Sparse signals only occupy a tiny fraction of the entire signal space and thus have a small amount of information, relative to their dimension. The theory tells us that the information can be captured faithfully with few random measurement samples, even far below the Nyquist rate. Despite the successful story, we question how the theory would change if we had a more precise prior than the simple sparsity model. Hence, we consider the settings where the prior is encoded as a probability density. In a Bayesian perspective, we see the signal recovery as an inference, in which we estimate the unmeasured dimensions of the signal given the incomplete measurements. We claim that good sensors should somehow be designed to minimize the uncertainty of the inference. In this thesis, we primarily use Shannon's entropy to measure the uncertainty and in effect pursue the InfoMax principle, rather than the restricted isometry property, in optimizing the sensors. By approximate analysis on sparse signals, we found random projections, typical in the compressed sensing literature, to be InfoMax optimal if the sparse coefficients are independent and identically distributed (i.i.d.). If not, however, we could find a different set of projections which, in signal reconstruction, consistently outperformed random or other types of measurements. In particular, if the coefficients are groupwise i.i.d., groupwise random projections with nonuniform sampling rate per group prove asymptotically Info- Max optimal. Such a groupwise i.i.d. pattern roughly appears in natural images when the wavelet basis is partitioned into groups according to the scale. Consequently, we applied the groupwise random projections to the sensing of natural images. We also considered designing an optimal color filter array for single-chip cameras. In this case, the feasible set of projections is highly restricted because multiplexing across pixels is not allowed. Nevertheless, our principle still applies. By minimizing the uncertainty of the unmeasured colors given the measured ones, we could find new color filter arrays which showed better demosaicking performance in comparison with Bayer or other existing color filter arrays.
by Hyun Sung Chang.
Ph.D.
156 p.
eng
Massachusetts Institute of Technology
M.I.T. theses are protected by
copyright. They may be viewed from this source for any purpose, but
reproduction or distribution in any format is prohibited without written
permission. See provided URL for inquiries about permission.
http://dspace.mit.edu/handle/1721.1/7582
Electrical Engineering and Computer Science.
Informative sensing : theory and applications
Thesis
application/pdf
Publication
unknown
unknown
<Publication xmlns="https://www.openaire.eu/cerif-profile/1.1/" id="5c3656f6-758b-405e-a3b3-8060ad6ece9f">
<Type xmlns="https://www.openaire.eu/cerif-profile/vocab/COAR_Publication_Types">http://purl.org/coar/resource_type/c_1843</Type>
<Language>eng</Language>
<Title>Informative sensing : theory and applications</Title>
<PublishedIn>
<Publication>
</Publication>
</PublishedIn>
<PublicationDate>2012</PublicationDate>
<Authors>
<Author>
<DisplayName>Chang, Hyun Sung</DisplayName>
<Affiliation>
<OrgUnit>
</OrgUnit>
</Affiliation>
</Author>
</Authors>
<Editors>
</Editors>
<Publishers>
<Publisher>
<DisplayName>Massachusetts Institute of Technology</DisplayName>
<OrgUnit />
</Publisher>
</Publishers>
<License>http://dspace.mit.edu/handle/1721.1/7582</License>
<Keyword>Electrical Engineering and Computer Science.</Keyword>
<Abstract>Compressed sensing is a recent theory for the sampling and reconstruction of sparse signals. Sparse signals only occupy a tiny fraction of the entire signal space and thus have a small amount of information, relative to their dimension. The theory tells us that the information can be captured faithfully with few random measurement samples, even far below the Nyquist rate. Despite the successful story, we question how the theory would change if we had a more precise prior than the simple sparsity model. Hence, we consider the settings where the prior is encoded as a probability density. In a Bayesian perspective, we see the signal recovery as an inference, in which we estimate the unmeasured dimensions of the signal given the incomplete measurements. We claim that good sensors should somehow be designed to minimize the uncertainty of the inference. In this thesis, we primarily use Shannon's entropy to measure the uncertainty and in effect pursue the InfoMax principle, rather than the restricted isometry property, in optimizing the sensors. By approximate analysis on sparse signals, we found random projections, typical in the compressed sensing literature, to be InfoMax optimal if the sparse coefficients are independent and identically distributed (i.i.d.). If not, however, we could find a different set of projections which, in signal reconstruction, consistently outperformed random or other types of measurements. In particular, if the coefficients are groupwise i.i.d., groupwise random projections with nonuniform sampling rate per group prove asymptotically Info- Max optimal. Such a groupwise i.i.d. pattern roughly appears in natural images when the wavelet basis is partitioned into groups according to the scale. Consequently, we applied the groupwise random projections to the sensing of natural images. We also considered designing an optimal color filter array for single-chip cameras. In this case, the feasible set of projections is highly restricted because multiplexing across pixels is not allowed. Nevertheless, our principle still applies. By minimizing the uncertainty of the unmeasured colors given the measured ones, we could find new color filter arrays which showed better demosaicking performance in comparison with Bayer or other existing color filter arrays.</Abstract>
<Access xmlns="http://purl.org/coar/access_right"
>
</Access>
</Publication>