On predictive modeling and classification of signals /
Most real world signals have the highly correlated nature between signal components. By utilizing this correlative structure of signals, one can obtain appropriate representations for a wide variety of applications. In this dissertation, we have studied a number of problems in image processing and t...
| Main Author: | |
|---|---|
| Format: | Thesis Book |
| Language: | English |
| Published: |
[Place of publication not identified] :
[publisher not identified] ;
1999.
|
| Subjects: | |
| Online Access: | http://proxy.library.tamu.edu/login?url=http://proquest.umi.com/pqdweb?did=730298401&sid=1&Fmt=2&clientId=2945&RQT=309&VName=PQD |
| Summary: | Most real world signals have the highly correlated nature between signal components. By utilizing this correlative structure of signals, one can obtain appropriate representations for a wide variety of applications. In this dissertation, we have studied a number of problems in image processing and time-series analysis based on the predictive modeling that captures the correlative structure over the regions of interests. As specific applications, we studied edge detection, texture segmentation, removal/reduction of impulse noise, and modular approach to time-series prediction. For the edge detection problem including tracking of texture boundaries, image/texture patterns round a pixel p are used to predict the value of p. If the variance of squared prediction errors is small, the region around p is considered as uniform region. By employing the encoding by feature map indices to this scheme, we could extend the edge detection to the texture boundary tracking. We approached the texture segmentation problem by combining the texture boundary tracking process with the class labeling process. For the removal of impulse noise, we showed that the decision-based median altering can be improved by accurately estimating if the center pixel of a window is Corrupted by noise. In our approach, the predictions from the neighboring pixels are used as decision criteria. We also showed that accuracy of time-series prediction can be improved by adding backward prediction module to the usual forward prediction Module. Finally, we developed a modular system for predicting the link travel time on a transportation network, in which pattern classification is performed by partitioning the input space according to prediction errors. |
|---|---|
| Item Description: | Vita. "Major Subject: Computer Science". |
| Physical Description: | xv, 162 leaves : illustrations ; 28 cm. Issued also on microfiche from University Microfilm Inc. |
| Bibliography: | Includes bibliographical references (leaves 145-161). |