A lightweight YOLOv8-based UAVs detection with dual-path optimization and attention mechanism

The rapid proliferation of unmanned aerial vehicles (UAVs) has introduced significant security challenges, including threats to public safety and personal privacy, due to their compact size and high maneuverability. Existing visual detection algorithms struggle with critical limitations in complex scenarios: frequent small-target misses and excessive computational demands. To address these issues, we propose an optimized lightweight detection incorporating two key innovations. We design the C2fDBSC lightweight module, which integrates depthwise separable convolution and bidirectional separation convolution to achieve efficient multi-scale feature representation while significantly reducing model parameters. This enables enhanced computational efficiency, making the model well-suited for deployment on edge devices. Complementing this, a Selective Kernel-enhanced attention mechanism counters complex background interference through channel-spatial dual-domain feature recalibration, where parallel heterogeneous convolutional branches adaptively capture rotor textures and flight trajectory patterns via gate-controlled feature fusion optimized for altitude variations. Comprehensive evaluations on the DUT Anti-UAV dataset under standard evaluation protocols demonstrate our model's superiority: achieving 91.7% [email protected] with merely 10.3 MB parameters, representing a 1.5% accuracy improvement over YOLOv8s baseline. The architecture maintains real-time detection capability at 77.3 frames per second (FPS), establishing a new balance between computational efficiency and detection precision for UAVs surveillance applications.