Ensembles of classifiers for improved SAR image recognition using pseudo Zernike moments

Abstract

In this paper, a new approach for improving the classification of different kinds of ground vehicles from moving stationary target acquisition and recognition images is proposed. Pseudo Zernike moments are used for feature extraction due to its capability of being scale, rotation, and translation invariant. To benefit from the diversities of regions we utilize both target and shadow regions as separate regions of interest for vehicle representation. Region of interests in the form of “area,”“boundary,” and “texture” are used for extraction. Extracted features from target and shadow regions of area, boundary, and texture are fused and fed to different classifiers. Five classifiers with different properties are adopted, including support vector machine, which is a parametric classifier that can control overfitting, in contrast to the decision tree, which is a nonparametric classifier, linear discriminant analysis, and k-nearest neighbor, which have cheaper computational cost, and random forest, which is an appropriate classifier for estimating outlier and missing data. In order to improve the overall performance of target recognition, we proposed a novel approach in which first we define six regions and fuse them to a single vector. Then fused feature vectors are fed to classifiers and the final decision is generated using majority voting. Experimental results justify that by combining decision with majority voting the performance is improved.