High-Order Feature Interaction Networks for Lightweight Real-Time Object Detection

Zhiqi Shao

Abstract: Real-time object detectors have achieved a strong balance between accuracy and efficiency, yet their feature interaction mechanisms remain largely constrained by local aggregation or pairwise attention. This limitation becomes evident in crowded scenes, scale-varying objects, and cluttered backgrounds, where detection requires reasoning over global relations among multiple spatial regions and semantic levels. In this paper, we present a lightweight YOLO-style detector built around high-order feature interaction. The core component is an adaptive hypergraph correlation module that models latent multi-to-multi dependencies among feature locations and scales, enabling efficient global information exchange without introducing heavy transformer-style computation. Based on this module, we further design a full-pipeline aggregation and redistribution strategy, allowing correlation-enhanced representations to flow through the backbone, neck, and detection head in a coordinated manner. To keep the model practical for real-time deployment, large-kernel operations are replaced with depthwise separable convolutional blocks, reducing parameters and FLOPs while preserving receptive-field coverage. Experiments on the MS COCO benchmark show that the proposed detector improves detection accuracy over YOLOv11-N, YOLO12-N and YOLOv13-N under comparable computational budgets. The results indicate that high-order correlation modeling can offer a favorable accuracy-efficiency trade-off for real-time object detection in complex visual environments.