SFD-YOLO for small-object fragment impact detection in warhead target-plate testing.

To address missed detections and limited efficiency in small fragment-impact recognition on target plates during warhead testing, this paper proposes small fragment detection YOLO (SFD-YOLO), a task-oriented detection framework built upon YOLOv11. An improved Spatial–Channel Reconstruction C3k2 (SCC3k2) module is integrated into the backbone to suppress redundant responses in both spatial and channel dimensions, enhancing the representation of weak micro-scale impact cues. To improve sensitivity to extremely small targets, we introduce an additional micro-object detection head and adopt an Asymptotic Feature Pyramid Network (AFPN) for progressive multi-level feature alignment and fusion, which strengthens feature consistency across pyramid levels. In addition, a Lightweight Adaptive Extraction (LAE) module is employed to replace standard convolutions, reducing model complexity while maintaining effective feature extraction. To comprehensively evaluate performance in realistic testing scenarios, we construct a multi-scene target-plate dataset from a series of static explosion experiments, covering both penetrative fragment holes and non-penetrative impact marks. Experimental results demonstrate that SFD-YOLO achieves 98.1% mAP@0.5 and 69.7% mAP@0.5:0.95, outperforming the YOLOv11 baseline by 2.7% in mAP@0.5 and 6.8% in mAP@0.5:0.95, at 135 FPS with only 2.15M parameters. Moreover, robustness evaluations under image degradations indicate that SFD-YOLO maintains more stable detection performance than the baseline. The proposed method provides a high-precision real-time solution for fragment lethality evaluation and shows potential for broader applications such as metal surface defect inspection.