Abstract
<title>Abstract</title> <p>A fundamental conflict in deploying lightweight object detectors on UAVs is that the downsampling for efficiency inevitably loses features of small targets, degrading detection. To resolve this, we propose RLPF-YOLO, guided by a “feature compensation” design philosophy to actively recover and enhance lost information. We design a novel backbone with three key modules: 1) a Dual-Branch Feature Complementary Module (C2F-FCM) for coarse and fine feature fusion; 2) an Efficient Multi-Scale Convolution Module (RMKPConv) to capture con-textual information with minimal cost; and 3) a Residual Attention Module (RCBAM) to dynamically emphasize crucial spatial and channel features for small objects. A Cross-Stage Partial Feature Processing Module (CSP-SimAM) is also introduced in the neck for efficient multi-scale fusion. Experiments on VisDrone2019 show that at 640-pixel resolution,RLPF-YOLO reduces the number of parameters by 70.4%, while increasing mAP@50 from 32.2% to 34.2% and mAP@(50-95) from 18.6% to 20.1%, respectively. Crucially, at a high 1280-pixel resolution,It also outperforms the baseline in terms of precision, demonstrating excellent scale robustness and practical deployment value.This work provides an effective architectural solution to unify the typically opposing goals of lightweight design and high-accuracy small-object detection.The code is available at:https://github.com/jsdghka/yolo-RLPF.git</p>