Effect of Partial Fine-Tuning of SqueezeNet on MSTAR for Automatic Military Target Recognition

Synthetic Aperture Radar (SAR) is widely studied for automatic target recognition (ATR) due to its operational independence from lighting conditions and long-range capabilities. The Moving and Stationary Target Acquisition and Recognition (MSTAR) dataset is a benchmark for the evaluation of ATR SAR techniques, such as deep neural networks. Transfer learning, exploiting pre-trained models such as VGG16 and AlexNet, has shown promising results in ATR SAR by tuning all network parameters. This study investigates the effectiveness of the compact SqueezeNet model for ATR SAR, analysing the degree of tuning required to achieve state-of-the-art accuracy and exploring model compression techniques. The results contribute to improving the efficiency and accuracy of ATR SAR applications.