Lightweight Deep Learning Model for SAR Image Recognition

Synthetic Aperture Radar (SAR) image classification is very important in modern remote sensing applications, such as military surveillance, disaster management, and autonomous vehicle guidance. While traditional deep learning models are accurate, they are generally computationally expensive and have many parameters, making them less applicable in resource-constrained environments. In this paper, we propose a new light-weight architecture based on capsule optimization for SAR target recognition on the MSTAR benchmark dataset, named CPLO-LWRCapsNet. The proposed pipeline involves guided filtering for speckle noise suppression, Discrete Hartley Transform (DHT) for feature enhancement in the frequency domain, KAN-Driven Multiscale Synergy Network (KDMSNet) for accurate target segmentation, and Relative Directional Edge Binary Pattern (RDEBP) descriptors for robust texture feature extraction. The architecture hyperparameters are tuned by the Capsule-based Pareto-Level Optimization (CPLO) algorithm. The experiment results show that the test accuracy is 99.72% with only 210,357 trainable parameters and 172.6 million MACs. This exemplifies the capacity of the model to discover a suitable trade-off between high classification performance and computational efficiency.