Geometric invariant Target classification using 2D Mellin cepstrum with modified grid formation

The Classification of Targets in Synthetic Aperture Radar Images is greatly affected by scale, rotation and translation. This paper proposes a geometric invariant algorithm to classify military targets based on extracting cepstral features derived from the modified grid selection over spectral components of Fourier Mellin Transform. The proposed non uniform grid is formed by a window with a cell of 2×2 pixels at the center, surrounded by the cells of 4×4 pixels, and so on, with overlapping concept to extract better representative features. Further each cell is divided into upper and lower triangular bins. The energy of each bin forms the down sampled M×M data accounting the larger value between the two triangles so that the information is enhanced. The experiments are carried out with a total of 700 SAR images collected from MSTAR database with different combinations of rotation, scale and translations. The proposed method has been tested against existing methods such as Region Covariance, Co-differencing and 2D Mellin cepstrum with non- overlapping grids. The results from 2D-Mellin Cepstrum using the proposed grid formation have been observed to be better in terms of 92% detection accuracy compared with 86% for region covariance method and 89% for non-uniform grid formation method.