High resolution synthetic aperture radar images usually contain much redundant, noisy and irrelevant
information. Eliminating these information or extracting only useful information can enhance ATR
performance, reduce processing time and increase the robustness of the ATR systems. Most existing
feature extraction methods are either computationally expensive or can only provide ad hoc solutions
and have no guarantee of optimality. In this paper, we describe a new distance metric learning algorithm.
The algorithm is based on the local learning strategy and is formulated as a convex optimization
problem. The algorithm not only is capable of learning the feature significance and feature correlations
in a high dimensional space but also is very easy to implement with guaranteed global optimality.
Experimental results based on the MSTAR database are presented to demonstrate the effectiveness of
the new algorithm.
A Bayesian network (BN) is a directed acyclic graphical model that encodes probabilistic relationships among variables of interest. BNs not only provide a natural and compact way to represent the domain knowledge and encode joint probability distributions, but also provide a basis for efficient probabilistic inference. We apply BNs to wide area airborne minefield detection (WAAMD) due to their powerful representation ability of encoding the domain knowledge and their flexible structural extendibility for multi-look and multi-sensor data fusion. We first design BN models for both single-look detection and multi-look and multi-sensor data fusion and then refine them via learning from data using a structural expectation-maximization (SEM) algorithm. We evaluate the performance of our landmine detection scheme using data sets collected by three airborne ground penetrating synthetic aperture radars (GPSARs) (Lynx Ku-band, Mirage stepped-frequency (0.3 - 2.8 GHz), and Veridian X-band GPSARs) from various testing sites that have different terrain and vegetation conditions. Experimental results indicate that BNs can help improve the landmine detection performance significantly. The use of BNs for multi-look and multi-sensor data fusion is also shown to provide significant false alarm reductions.
Landmine detection using radar is a very challenging problem due
to weak signal returns of landmines and extremely complicated
surveying environments. In this paper, we present a new landmine
detection system using forward-looking ground penetrating radar
(FLGPR), which has shown a promising result in a recently
conducted blind test. The system uses wavelet packet transform and
the sequential feature selection algorithm to extract the most
discriminant information distributed in the joint time-frequency
domain for detecting landmines. We also propose a cascade training
method that allows a WPT based detector to continue learning from
the errors made on the unseen environment to improve its detection
performance. The effectiveness of the proposed detector is
demonstrated through a blind test based on the measured FLGPR data
collected over an area of 14400 square meters.
We propose a novel automatic target recognition (ATR) system for classification of three types of ground vehicles in the MSTAR public release database. First, each image chip is pre-processed by extracting fine and raw feature sets, where raw features compensate for the target pose estimation error that corrupts fine image features. Then, the chips are classified by using the adaptive boosting (AdaBoost) algorithm with the radial basis function (RBF) net as the base learner. Since the RBF net is a binary classifier, we decompose our multiclass problem into a set of binary ones through the
error-correcting output codes (ECOC) method, specifying a dictionary of code words for the set of three possible classes. AdaBoost combines the classification results of the RBF net for each binary problem into a code word, which is then "decoded" as one of the code words (i.e., ground-vehicle classes) in the specified dictionary. Along with classification, within the AdaBoost framework, we also conduct efficient fusion of the fine and raw image-feature vectors. The results of large-scale experiments demonstrate that our ATR scheme outperforms the state-of-the-art systems reported in the literature.
We consider landmine detection using forward-looking ground penetrating radar (FLGPR), which is quite challenging due to the weak signal returns of landmines. The two main challenging tasks include
extracting intricate structures of the target signals from the
radar imagery and adapting the classifier to the surrounding environment through learning. Through the time-frequency analysis, we find that the most discriminant information is time-frequency
localized. This observation motivates us to use the wavelet packet transform to sparsely represent the signals with the discriminant information encoded into several bases. Then the sequential floating forward selection method is used to extract these components and thereby a neural network classifier is designed. To further improve the classification performance, the AdaBoost algorithm is used. We modify the original AdaBoost algorithm to integrate the feature selection process into each iteration. Experimental results based on measured FLGPR data are presented, showing that with the proposed classifier, a significant improvement on both the training and the testing performances can be achieved.
We use the time-frequency analysis techniques for buried plastic landmine detection with a forward-looking Ground Penetrating Radar (GPR) system. Several time-frequency distributions are considered to
characterize and interpret the scattering phenomena of both targets and clutter. An ambiguity function based detector is also proposed, which employs principal component analysis for data dimensionality reduction and linear discriminant analysis for feature selection. Experimental results based on the SRI (Stanford Research Institute) experimentally measured forward-looking GPR data are presented, showing a significant detection performance improvement over the conventional detector.