PFC-MS: A Peripheral-Foveal Multi-Scale Network with Star Operation for Vehicle Detection

Abstract

To address the challenges of drastic scale variation, dense distribution, and illumination interference in vehicle detection from drone imagery, this paper proposes the PFC-MS framework. Built upon YOLO11n-obb, the framework integrates three modules: StarBlock leverages the star operation to map features into an implicit high-dimensional space, thereby enhancing feature representation; the Peripheral-Foveal Convolution (PFC) module, inspired by human vision, simulates the “sweep first, scrutinize later” mechanism to improve perceptual capability; and Multi-Scale Grouped Dilated Convolution (MSGDC) captures short-, medium-, and long-range dependencies through parallel multi-branch operations, strengthening multi-scale fusion in the neck. On the DroneVehicle dataset, PFC-MS achieves 77.2% mAP50 and 61.1% mAP50:95.