Compute Once, Distribute Many: Efficient Feature Pyramid Refinement for Ultra-Lightweight Detectors

The YOLO series has achieved remarkable progress in real-time object detection, yet nano-scale variants remain constrained by single-round feature pyramid topologies where the global correlation prior is distributed only once and extreme pyramid levels exchange information indirectly. In this paper, we propose YOLO-Prism, an efficient nano-scale detector built upon YOLOv13 that addresses this single-round limitation through asymmetric multi-round refinement. We propose a Multi-Round Feature Pyramid Network (MR-FPN) that exploits the decoupling between YOLOv13’s precomputed correlation prior and its lightweight gated distribution mechanism, appending a secondary top-down refinement pass at only ∼6% parameter overhead under a “compute once, distribute many” paradigm. Subsequently, as a near-zero-cost complement, we selectively integrate DCT-based Multi-Spectral Channel Attention (MSCA) at gradient-identified critical layers, contributing an additional gain at merely 0.01M parameters. Extensive experiments on MS COCO, PAS-CAL VOC, and VisDrone benchmarks demonstrate that YOLO-xsPrism achieves 41.66% AP50:95 on COCO val2017, surpassing the strong YOLOv13n baseline by 0.62 points under identical training conditions and outperforming YOLOv12n by 1.10 points, at 2.63M parameters and 7.1 GFLOPs. On YOLOv11n, whose neck lacks such a prior, the same asymmetric topology yields 0.50 AP50:95 at 3.8% overhead, confirming that the benefit of multi-round refinement is independent of any specific gating mechanism. This lightweight algorithm can be applied to intelligent video measurement, industrial edge detection sensors, and IoT monitoring.