SaKD: Lightweight infrared small target detection via Spatial-Aware Knowledge Distillation for edge deployment

Accurate infrared small target detection (IRSTD) in remote sensing remains challenging because targets of interest typically occupy only a few pixels and provide limited spatial cues against complex backgrounds. Existing high-performance methods rely on complex networks to enhance weak target representations, limiting their use on resource-constrained edge devices. Lightweight models reduce computation but struggle to capture scarce spatial cues of small targets. To preserve these spatial cues in lightweight models for remote sensing applications, we propose SaKD, a Spatial-Aware Knowledge Distillation method for IRSTD. SaKD explicitly captures complementary spatial cues associated with target responses from teacher predictions, thereby providing richer spatial discriminative supervision for the student model. Without changing the student inference structure, SaKD introduces two training stage components. The Multi-Scale Self Guidance Mechanism first derives auxiliary supervision from the student’s own responses at multiple scales to strengthen target representation before teacher guidance is introduced. Spatial Prior Distillation then transfers the teacher’s spatial priors by enforcing region, position, and scale consistency, corresponding to target region coverage, center localization, and response scale control. Experiments on multiple IRSTD datasets containing real remote sensing imagery show that SaKD consistently improves the undistilled lightweight student and achieves competitive detection performance with low model complexity. The distilled student contains only 0.895M parameters and 1.999G FLOPs. Deployment verification on a Jetson platform further demonstrates its practical efficiency for edge deployable remote sensing applications, reaching 142.7 FPS for standalone TensorRT FP16 inference at an input resolution of 256 × 256. The code will be available at: https://github.com/Kingwin97/SaKD .