Combined Deep learning architectures using Fusion Rules for Target Recognition in SAR images

In radar image detection and classification systems, one of the main challenges lies in the analysis and interpretation of captured images, whether for moving or stationary targets. This challenge has become increasingly complex due to the large volumes of radar data processed. Consequently, various techniques and methods for radar target detection and classification have been adopted, particularly those utilizing artificial intelligence paradigms. In this context, we explore data fusion techniques for Synthetic Aperture Radar (SAR) image classification. We propose a novel parallel architecture featuring three classifiers: the basic CNN, Inception-v3, and ResNet 50. The elementary decisions from each classifier are sent to a fusion center for the final decision. At the fusion center, a majority voting system is applied to consolidate the decisions from these classifiers. To evaluate the performance of this approach, we use the MSTAR (Moving and Stationnary Target Acquisition and Recognition) dataset. Simulation results demonstrate the effectiveness of the proposed approach, achieving a correct classification rate of 99.23%, which surpasses the performance of each individual CNN model.