H-SGE: A hybrid model based on scene graph enrichment for automated Handgun detection in security surveillance

Abstract

Small handgun detection in CCTV surveillance suffers from high false positives and negatives due to limited distinguishing features. We propose H-SGE (Hybrid Scene Graph Enrichment), combining Generative Adversarial Networks (GANs), scene graph enrichment, and multiple YOLO variants (YOLOv5, YOLOv7, YOLO10, YOLO11) for enhanced detection. H-SGE employs a five-stage pipeline: (1) enriched scene graph generation, (2) GAN-based feature enhancement, (3) context-aware RoI selection, (4) multi-YOLO detection, and (5) output fusion. Evaluation on a handgun dataset demonstrates significant F1-score improvements: YOLOv5 from 58% to 80%, YOLOv7 from 56% to 82%, YOLO10 from 62% to 85%, and YOLO11 from 64% to 87%, achieving over 20% accuracy gains. Results show that combining contextual reasoning with visual augmentation and hybrid detection effectively addresses small object detection challenges in surveillance systems.