Weapon detection in surveillance imagery is a critical task for ensuring public safety in high-risk environments. While single-stage object detectors like YOLOv8 provide exceptional localization speed, their semantic reliability often diminishes when distinguishing between lethal weapons and visually similar non-threat objects in complex scenes. To address this limitation, this work proposes a hybrid weapon detection framework that decouples object localization from semantic verification. The system utilizes a YOLOv8n detector for real-time region localization, followed by an attention-enhanced classification stage incorporating a Squeeze-and-Excitation (SE) mechanism. A novel conditional decision logic was implemented to manage the consensus between the two stages, incorporating a confidence-weighted fail-safe to maintain high recall for unambiguous threats. We evaluated three backbone architectures—NASNetMobile, DenseNet201, and InceptionV3, under identical constraints. Experimental results on a primary test set show that the hybrid NASNetMobile + SE model achieves 96% accuracy, significantly outperforming the standalone YOLOv8n (94%). More critically, on a challenging ablation dataset, the hybrid framework improved pistol precision from 90% to 95%. Furthermore, the integration of the SE block was found to optimize inference, reducing latency compared to the vanilla backbone. The proposed system represents a “Pareto-optimal” solution for edge-deployed security infrastructure, prioritizing high-precision firearm recognition without compromising operational latency.