Lightweight Deep Learning Model for SAR Image Recognition: Survey

SAR imagery can perform all-weather imaging in the defence and environmental monitoring arena. Although deep learning is a turning point in image recognition, its application in SAR ATR is presently thwarted by major technical bottlenecks. The state-of-the-art models presented in these papers are mostly based on heavy Convolutional Neural Networks (CNNs) that were originally designed for optical imagery. The high parameter counts and memory usage of these architectures make them unsuited for resource-constrained edge devices such as UAVs. Training complex models on scarce labelled SAR datasets also tends to result in overfitting. To deal with these challenges, there is an urgent need for both efficiency and scalability to meet future demand. We introduce the AutoSAR framework, an approach that unifies lightweight architectures (e.g., MobileNetV3), Neural Architecture Search (NAS), and modern model compression techniques such as pruning and quantization. This work synthesizes the current state of the art in the field and provides a starting point for model development, with the aim of maximizing size reduction while maintaining a high level of recognition accuracy for real-time processing of SAR data on operational platforms.