Related Experiment Videos
Use of multiresolution wavelet feature pyramids for automatic registration of multisensor imagery
Ilya Zavorin1, Jacqueline Le Moigne
1Goddard Earth Science and Technology Center, University of Maryland Baltimore County, Catonsville, MD 21228, USA. zavorin@backserv.gsfc.nasa.gov
Summary
This study evaluates wavelet pyramids for image registration accuracy and speed. Bandpass wavelets from the steerable pyramid offer the best accuracy, while low-pass wavelets improve convergence for satellite imagery registration.
Area of Science:
- Digital imaging
- Remote sensing
- Computer vision
- Medical imaging
Background:
- Image registration is crucial across disciplines using digital imaging.
- Developing robust, fast, and accurate automatic registration algorithms is essential for various imaging platforms.
- Selecting appropriate image information for geometric transformation search is a key challenge.
Purpose of the Study:
- To evaluate wavelet pyramids for invariant feature extraction and multi-resolution image representation.
- To accelerate image registration processes.
- To identify optimal wavelet features for robust image alignment.
Main Methods:
- Evaluation of several wavelet pyramids for image registration.
- Utilizing steerable pyramids (Simoncelli) for feature extraction.
- Modifying a gradient-based registration algorithm for satellite imagery.
Main Results:
- Bandpass wavelets from the steerable pyramid demonstrated superior accuracy and consistency.
- Low-pass wavelets from the same pyramid yielded the best radius of convergence.
- The modified gradient-based algorithm showed effectiveness on real and synthetic satellite data.
Conclusions:
- Wavelet pyramids, particularly steerable ones, are effective tools for image registration.
- The choice between bandpass and low-pass wavelets impacts accuracy versus convergence.
- The proposed modified algorithm enhances satellite image registration capabilities.