Spot SAR ATR Using Wavelet Features and Neural Network Classifier
Abstract
An overview and performance summary of an Automated Target Recognition (ATR) algorithm based on spot Synthetic Aperture Radar (SAR) imagery is described in this report. Feature extraction and classification are very important steps in the ATR process. In this algorithm, the two dimensional wavelet decomposition method was applied to SAR targets to extract features. Selection of an appropriate mother wavelet was done by testing various wavelets and selecting the one which produced the smallest variation between features for the same target types, and the largest variation between features for different target types. After extensive testing, the Reverse Biorthogonal was selected as the best mother wavelet for this application. Second level approximation coefficients were used as features. and were fed into a Multi Layer Perceptron (MLP) neural network (NN) for classification. The MLP MN was trained using a supervised method, the standard delta rule. The classification results are shown using Receiver Operation Characteristic (ROC) curves and Confusion Matrices. The analysed result shows that the Reverse biorthogonal wavelet features are as good as two-dimensional Fast Fourier Transform features in the MSTAR (Moving and Stationary Target Acquisition and Recognition) dataset application. Results also show that including confusers (objects that the ATR algorithm is not intended to classify) in the training dataset reduces false alarm because the classifier has learned to reject confusers during the training process.
Document Details
- Document Type
- Technical Report
- Publication Date
- Oct 01, 2005
- Accession Number
- ADA443104
Entities
People
- N. M. Sandirasegaram
Organizations
- Defence Research and Development Canada