Beyond Convolutions YOLOv8 with DRFF-Based Channel-Mixing for Enhanced Drone Detection

Detecting small unmanned aerial vehicles (UAVs) using standard object detectors remains challenging due to their limited ability to incorporate global contextual information and their reliance on localized convolutional operations. To address this limitation, this paper proposes an enhanced YOLOv8-based detection framework by integrating a Deep Residual Feed-Forward Network (DRFF) into the neck architecture. Inspired by the feed-forward blocks of transformer models, the DRFF introduces dense global channel-wise mixing at multi-scale feature fusion stages, enabling more effective integration of fine-grained spatial details and global contextual cues. The proposed YOLOv8-DRFF model is evaluated on a publicly available multi-view drone tracking dataset captured under real-world outdoor conditions using heterogeneous camera setups. Experimental results demonstrate that the proposed method significantly outperforms a fine-tuned YOLOv8 baseline, achieving an F1-score of 0.9567, an average IoU of 0.8270, and a mAP of 0.982, corresponding to improvements of 8.5% in F1-score and 6.7% in IoU. Despite a modest increase in computational complexity, the model maintains real-time inference performance at approximately 60 FPS. These results confirm that the proposed DRFF-enhanced YOLOv8 provides an effective and reliable solution for robust small drone detection in complex and cluttered environments.