Analysis of Deep Neural Networks for Military Target Classification using Synthetic Aperture Radar Images

Target detection and classification in the military is an area that is very significant in modern battlefields. Using Synthetic Aperture Radar images for classifying targets adds to its significance, as these images are high-resolution images of the surface of the earth created using microwave radiation and they can be used anytime, anywhere, and in any weather conditions. A target classification system using deep learning to classify military vehicles from Synthetic Aperture Radar images is proposed in this study. The system uses a baseline Convolutional Neural Network to classify the images of military vehicles from the MSTAR dataset, achieving a baseline accuracy of 90%. Further transfer learning was applied to the system by using 5 different pre-trained networks, namely the InceptionV3, VGG16, VGG19, ResNet50, and MobileNet. These models were analysed and evaluated using 3 different evaluation metrics, the Confusion matrix, Classification report, and Mean Average Precision to discover the most accurate and efficient model for this task. The models VGG16 and MobileNet displayed the best performance on the dataset achieving accuracies of 98% and 97%, respectively. The ResNet50 model displayed the worst performance among the models, achieving an accuracy of 82%. While the other models, InceptionV3 and VGG19, achieved accuracies of 92% and 96% respectively.