SFBF-YOLO for Small Object Detection in UAV and Remote Sensing Images

Small object detection (SOD) in UAV imagery is fundamentally constrained by the degradation of fine-grained information during hierarchical feature extraction, where conventional downsampling operations induce an inherent representation conflict between high-resolution spatial details and high-level semantic abstraction. To address this, we propose SFBF-YOLO, an algorithm centered on a novel dual-stream reconstruction mechanism. Instantiated by the space fine-grained attention module (SFGA), this mechanism serves as the architecture’s core performance driver. It bifurcates the information flow into a granularity stream that preserves subtle structural attributes, and a contextual stream that models global environmental dependencies, thereby reconstructing a robust feature manifold at the source. Building upon this foundation, we introduce balanced-PAN (BPAN) to enforce a symmetric feature equilibrium principle (SFEP). BPAN explicitly models the bidirectional reciprocity between classification-oriented semantics and localization-oriented precision across scales. Within this synergistic pipeline, a fast ghost-based semantic fusion module (FGSF) executes redundancy-aware feature evolution, ensuring high computational efficiency. Extensive evaluations on VisDrone-2019, USOD, and VEDAI benchmarks demonstrate that SFBF-YOLO achieves a superior trade-off between precision and efficiency. Notably, on VisDrone-2019, SFBF-YOLO-s achieves a 87% and 47% reduction in parameters and GFLOPs compared to YOLOv8m, while improving mAP₅₀ and mAP₅₀₋₉₅ by 0.044 and 0.029, respectively. Similarly, SFBF-YOLO-m outperforms YOLOv8l with 79% and 39% lower complexity, alongside accuracy gains of 0.049 and 0.028. These results empirically validate that the proposed dual-stream reconstruction mechanism and SFEP are pivotal for robust detection in complex remote sensing scenarios.