Recognition for Targets Type of Synthetic Aperture Radar Imagery using Deep Learning Methods

Abstract

In recent years, Synthetic Aperture Radar (SAR) imagery has played a crucial role in various applications such as reconnaissance, surveillance, and target recognition. Deep learning methods have emerged as powerful tools for target recognition in SAR imagery due to their ability to automatically learn discriminative features from data. We review state-of-the-art techniques for dataset preparation, network architecture design, training strategies, and performance evaluation in SAR target recognition tasks. Achieving a verification accuracy of around 97% for an eight-class classifier is remarkably impressive. In addition, the similarity between training and validation accuracy indicates the strength of our classifier. Overfitting, which has much higher training accuracy compared to validation accuracy, is a concern The depicted image outlines the training process. In the upper plot, the dark blue line represents the model's accuracy on the training data, while the black dashed line illustrates its accuracy on the separate validation data. Achieving a validation accuracy of nearly 97% for an eight-class classifier is notably impressive. Additionally, the similarity between the training and validation accuracies indicates the robustness of our classifier. The ZSU-23/4 class appears to present the most difficulty for the model among the eight classes. Due to similarities in SAR images, instances of misclassification occur between the ZSU-23/4 and 2S1 classes. Nonetheless, the model demonstrates the capability to achieve accuracy levels exceeding 90% for this class.