Super-Resolution Approach for Low-Cost and Low-Resolution Radars

Low-cost radars have emerged as attractive alternatives to optical sensors for consumer applications due to their robustness and inherent privacy. However, their limited spatial resolution constrains their applicability to tasks requiring fine motion analysis. This work proposes a deep learning-based super-resolution (SR) framework that enhances radar range- Doppler maps with task-specific fidelity. The proposed approach integrates a spatially adaptive feature modulation network and introduces classification-guided training that selectively refines gesture-relevant features. Unlike conventional image SR techniques focused on perceptual quality, the proposed approach jointly optimizes SR and classification objectives to impr ve radar-based hand gesture recognition (HGR). Three scalable architectures, SuperGestNet, MultiSuperGestNet, and RecursiveSuperGestNet, are proposed, offering trade-offs between accuracy, computational complexity, and memory. Experiments on real radar data demonstrate substantial improvements in HGR accuracy over conventional SR methods and baseline classifiers, even surpassing high-resolution models in some settings. This taskaware SR paradigm enables the practical deployment of low-cost, low-resolution radars for various consumer applications.