RMD-YOLO11n A Small Object Detection and Multi-scale Fusion Optimization Method for UAV Images

Abstract

To address the challenges of small object detection in UAV (Unmanned Aerial Vehicle) imagery, such as minute scales, complex backgrounds, and severe occlusions, this paper proposes an improved model based on YOLOv11n, designated as RMD-YOLO11n. First, the C3k2-RFAConv module is integrated into the backbone network, leveraging the Receptive Field Attention (RFA) mechanism to overcome the limitations of feature representation caused by parameter sharing; this effectively enhances the extraction of tiny object features while suppressing background noise. Second, the neck network is reconstructed by designing a GUFPN (Multi-branch Auxiliary Feature Pyramid Network) structure. Through the CUF and AUF modules, the retention of high-resolution spatial information is strengthened, compensating for the loss of target details during the downsampling process. Furthermore, the C2PSA-DYT module is introduced to optimize the feature attention mechanism, further improving training stability and feature focusing capabilities. Experimental results on the VisDrone2019 dataset demonstrate that RMD-YOLO11n achieves an mAP@0.5 of 35.8%, representing a 2.9% improvement over the baseline model. While maintaining its lightweight advantage with only 2.7M parameters, the model significantly enhances detection performance for multi-scale objects in complex scenarios. Confusion matrix analysis indicates that the proposed algorithm effectively reduces the missed detection rate in cluttered backgrounds.Specifically, the recall rates for tiny objects such as "Pedestrian" and "People" are increased by 4% and 3%, respectively, effectively mitigating the issue of targets being overwhelmed by the background in high-altitude UAV perspectives.