MCFNet: Lightweight Tiny-Object Detection for Edge Deployment

In long-range unmanned aerial vehicle (UAV) imagery, distant targets may occupy only a few pixels, making their spatial cues vulnerable to downsampling and background interference. MCFNet combines a selective P2/P4 prediction pathway with Cross Spatial-Channel Fusion (CSCF), Multi-Kernel Feature Fusion (FFM), and Adaptive Masking Attention (AMA). On a self-constructed aerial armor dataset, it achieves a mean average precision of 0.623 over intersection-over-union thresholds of 0.50–0.95 (mAP50-95), with 4.26 million parameters and 20.0 billion floating-point operations (FLOPs). On VisDrone-2019, mean mAP50 is 0.414, compared with 0.392 for YOLO11-S, with 54.7% fewer parameters. Representative scale analysis shows gains of 1.46 and 0.98 percentage points for Tiny and Small objects, respectively, while Non-Tiny mAP50 remains 0.413. On AI-TOD-v2, MCFNet increases mAP50-95 from 0.1417 to 0.1555 with 54.8% fewer parameters than YOLO11-S. RDK X5 8-bit integer (INT8) evaluation compares the models at matched resolutions and examines resolution adjustment within MCFNet. At 640 × 640, MCFNet achieves an mAP50-95 of 0.5305 at a forward latency of 22.95 ms, compared with 0.4457 at 15.75 ms for YOLO11-S. Reusing the MCFNet weights at 512 × 512 reduces latency to 13.74 ms while retaining mAP50-95 of 0.4628. Input resolution can therefore be adjusted to balance the accuracy and latency on the target hardware.