C3F-YOLO: A Contextual, Contrastive, and Compact Framework for Small Object Detection in UAV Imagery

Small object detection in Unmanned Aerial Vehicle (UAV) imagery remains a formidable challenge due to the inherent limitations of low pixel counts and weak feature representations. These factors are further exacerbated by complex backgrounds and the stringent computational constraints of UAV platforms. To address these issues, this paper introduces MCB-YOLO, a novel and efficient detection framework built upon the YOLO architecture. We systematically identify and tackle three key problems: firstly, the inability of standard convolutions to capture sufficient contextual information for minuscule objects; secondly, the model's susceptibility to distraction from cluttered backgrounds; and thirdly, the incompatibility of large model footprints with low-end embedded devices. Correspondingly, our solution incorporates three innovative components: (1) The Multi-Scale Context Fusion (MCF) module, designed to aggregate contextual information from diverse receptive fields, thereby enriching the feature representation of small targets. (2) The Spatial & Channel Dual Attention (SCDA) mechanism, which directs the model's focus towards salient object features while suppressing irrelevant background noise. (3) A re-parameterized lightweight backbone, utilizing cross-stage partial connections and depthwise convolutions to drastically reduce parameters and computational cost without sacrificing performance. Extensive experiments on the VisDrone2019 and DIOR datasets demonstrate that MCB-YOLO achieves a superior balance between accuracy and efficiency. Our model outperforms several state-of-the-art YOLO baselines, showing significant improvements in mean Average Precision (mAP) for small objects while maintaining a compact model size suitable for real-time deployment on UAV platforms.