Autonomous Time-frequency Cropping and Feature-extraction Algorithms for Classification of LPI Radar Modulations

Download or read book Autonomous Time-frequency Cropping and Feature-extraction Algorithms for Classification of LPI Radar Modulations written by and published by . This book was released on 2006 with total page 78 pages. Available in PDF, EPUB and Kindle. Book excerpt: Three autonomous cropping and feature extraction algorithms are examined that can be used for classification of low probability of intercept radar modulations using time-frequency (T-F) images. The first approach, Erosion Dilation Adaptive Binarization (EDAB), uses erosion and a new adaptive threshold binarization algorithm embedded within a recursive dilation process to determine the modulation energy centroid (radar's carrier frequency) and properly place a fixed-width cropping window. The second approach, Marginal Frequency Adaptive Binarization (MFAB), uses the marginal frequency distribution and the adaptive threshold binarization algorithm to determine the start and stop frequencies of the modulation energy to locate and adapt the size of the cropping window. The third approach, Fast Image Filtering, uses the fast Fourier transform and a Gaussian lowpass filter to isolate the modulation energy. The modulation is then cropped from the original T-F image and the adaptive binarization algorithm is used again to compute a binary feature vector for input into a classification network. The binary feature vector allows the image detail to be preserved without overwhelming the classification network that follows. A multi-layer perceptron and a radial basis function network are used for classification and the results are compared. Classification results for nine simulated radar modulations are shown to demonstrate the three feature-extraction approaches and quantify the performance of the algorithms. It is shown that the best results are obtained using the Choi-Williams distribution followed by the MFAB algorithm and a multi-layer perceptron. This setup produced an overall percent correct classification (Pcc) of 87.2% for testing with noise variation and 77.8% for testing with modulation variation. In an operational context, the ability to process and classify LPI signals autonomously allows the operator in the field to receive real-time results.