A neural network using wavelet for stereo vision research /
| Main Author: | |
|---|---|
| Other Authors: | , , |
| Format: | Thesis Book |
| Language: | English |
| Published: |
1994.
|
| Subjects: | |
| Online Access: | Link to OAKTrust copy ProQuest, Abstract |
| Abstract: | The purpose of stereo machine vision is to determine the three-dimensional location of objects from two two-dimensional images captured from different locations. The most difficult task in stereo vision applications is the determination of matching features in the two image planes, called correspondence problem. This dissertation concentrates on the utilization of the wavelet transform and neural network architecture jointly for improving this matching capability. The wavelet transform is used to facilitate a coarse-to-fine matching approach by using its Multi Resolution Analysis (MRA) properties and also to shift a set of candidate features along the epipolar lines within limited space. The wavelet transform coefficients are used in the stereo neural network, which is a hierarchical structure accommodating the coarse-to-fine matching strategy. The artificial neural network is introduced to achieve good performance through dense interconnection of non linear simple computational elements. The fusion mechanism is the process of minimizing the global Root Mean Square Error (RMSE) of the two signals with shifting operations, which are guided by the local matching criteria. The advantage of this network is that there is no multiple matching (only one disparity value), the performance does not depend on early processing, dense depth information is obtained rather than sparse depth information, and the method lends itself to massive parallel computing architecture. |
|---|---|
| Item Description: | Vita. "Major subject: Electrical Engineering." |
| Physical Description: | xi, 91 leaves : illustrations ; 28 cm |
| Bibliography: | Includes bibliographical references. |