SF-YOLO: a lightweight and precise approach for small object detection in aerial images

Detecting small objects in aerial images presents a persistent challenge due to their dense distribution, limited feature representation, and large-scale variations. For example, detecting small objects is harder for aerial data than for standard object detection tasks. Traditional object detection techniques struggle with maintaining detailed features, achieving balanced feature aggregation, and meeting computation requirements for low-end devices. YOLO-based models dominate real-time object detection due to their speed and efficiency, making them ideal for resource-constrained applications like UAVs and embedded systems. However, their performance degrades for small objects a critical limitation in aerial imaging, where small, densely clustered objects and large resolution variability distort bounding box distributions. This weakness stems from YOLO’s architectural biases toward larger objects and fixed grid priors, which poorly adapt to the scale and density shifts typical in aerial perspectives. To avoid these constraints, we introduce a lightweight design integrating RepVGGBlock, Selective Fusion Block (SF Block), Spatial Pyramid Pooling Fast with Width and Depth (inferred) (SPPF-WD) module, and Specialized Small-Object Detection Branch. This increases feature extraction for multi-scale feature fusion and small-object detection while maintaining a reduced architecture, which is beneficial in real-time scenarios on limited computational capacity platforms. The proposed model, SF-x, performs better on the SIMD dataset than the best existing models, with mAP@50 and mAP@50-95 that are 1.2% and 3.6% higher, respectively. Furthermore, it demonstrates efficacy by significantly improving performance at every object scale, thereby solidifying its state-of-the-art status in precision and robustness. The model shows a 14.8% increase in APS and a 10.5% reduction in false positives compared to baseline models. Importantly, the lightweight SF-n configuration (2.99M, 8.7 GFLOPs) is designed for edge devices. On the other hand, the scalable configurations (SF-n, SF-m, SF-l, SF-x) can achieve the best trade-off between accuracy, computational cost, and inference speed. The findings highlight the efficiency and adaptability of the proposed lightweight paradigm, paving the way for it to become a yardstick solution for small object detection in aerial and remote sensing images. Compared to similar state-of-the-art solutions, if you prioritize accuracy and real-time, you can use it in UAV surveillance, traffic, environmental assessment, or anything similar in a resource-constrained environment. The code and trained weights for the proposed framework can be found at this link: GitHub Repository: SF YOLO (https://github.com/mrshafayet/sfyolo)