Multiscale Anomaly Detection and Image Registration Algorithms for Airborne Landmine Detection
Abstract
Multiscale techniques such as the wavelet transform have become a powerful tool for image compression, denoising, and analysis. This thesis outlines the use of such techniques for developing a multiscale processing stream for the purpose of airborne landmine detection. In this work, the Reed-Xiaoli (RX) multispectral anomaly detection algorithm is extended to perform detection within the shift-invariant wavelet representation of a given single-band image. A test statistic is derived and shown to better detect anomalies in a correlated noise background than the single-band RX algorithm. The results of a detection algorithm using the shift-invariant wavelet representation and a multi-layer perceptron neural network are also discussed. A scale-space image registration algorithm is presented, where the scale-invariant feature transform (SIFT) has been used to search for control points between two images. The use of SIFT allows for image features invariant to scaling, translation and rotation to be matched in feature space, resulting in more accurate image registration than traditional correlation-based techniques.
Document Details
- Document Type
- Technical Report
- Publication Date
- May 01, 2008
- Accession Number
- ADA539722
Entities
People
- Jeffrey D. Barnes