Impact of Regularization in Convolutional Neural Networks for Military Target Classification Using SAR Images

In modern warfare, the military field of target recognition and classification is important.Using Synthetic Aperture Radar (SAR) images for classifying targets, especially ground military targets, brings an advantage since these are high-resolution images of the surface of the earth produced by microwave radiation and may be used anywhere, anytime and in all-weather conditions.Whereas SAR offers invaluable capabilities, speckle noise may impact the results.Instead of using despeckling filters, this study investigates the impact of two custom Convolutional Neural Networks (CNNs), a baseline architecture with no regularization techniques and a model with regularization techniques, which ensures an invariance to this type of noise.Using the Moving and Stationary Target Acquisition and Recognition (MSTAR) dataset to classify the images of military vehicles, the baseline architecture achieved an overall accuracy of 91% and the enhanced architecture achieved an overall accuracy of 99%.This paper shows that regularization techniques help with a more accurate classification, which is essential in the military domain.However, the enhanced architecture is sensitive when it comes to artificially added speckle noise.When speckle noise with a standard deviation of 0.45 is introduced, the overall accuracy decreases to 37% for the enhanced architecture, but the baseline architecture remains with an overall accuracy of around 84%.