YOLOv11-EASFF: an improved object detection model with early adaptive spatial feature fusion and structured pruning for railway foreign object intrusion detection

Abstract

Foreign object intrusion on railways represents a major hazard threatening operational safety. The characteristics of long distance and small size lead to high missed detection rates and poor scene adaptability in existing general-purpose object detection models. To address this, this paper proposes a railway foreign object intrusion detection model based on YOLOv11-EASFF. The model is built upon the lightweight version YOLOv11n and incorporates targeted improvements in three dimensions: feature fusion, scene adaptation, and model pruning. First, tailored to the severe scale variations and weak features of small targets in railway scenarios, an Early Adaptive Spatial Feature Fusion (EASFF) module is designed, which introduces an input-dependent dynamic two-level weight learning mechanism between the backbone network and the feature pyramid network. This mechanism adaptively enhances the fusion of multi-scale early features, especially details of small objects. Second, the anchor configuration is customized and optimized based on the size characteristics of railway foreign objects to improve the model’s scene adaptation. Finally, which for the first time synergistically integrates a module-level hard upper-bound protection mechanism with channel-level pruning based on BatchNorm γ statistics. Experiments show that even under a highly saturated baseline achieving a mAP@0.5 of 0.992, while achieving a parameter sparsity of 32.66% through pruning, It providing an efficient, reliable, and fast-processing technical solution for real-time warning of railway foreign object intrusion on edge devices.