Answers are grounded in this paper and cite their sources.
Detecting small objects in remote sensing images remains a challenging task due to the submergence of weak features in complex backgrounds, which is even harder under limited computational and storage resources. Existing detection algorithms often struggle to achieve an optimal trade-off between lightweight deployment and high detection precision. To address this issue, this article proposes GRO-YOLO, a lightweight detector based on a global-region-object multi-granularity perception architecture. First, Asymmetric Fusion Neck (AF-Neck) is designed to maximize the retention of spatial texture information for small objects, simultaneously, reduce computational redundancy. Second, to achieve high-precision detection under lightweight constraints, we construct a Magnitude-Aware Global Context (MAGC) module and a Dense Context Block (DCBlock). The former captures long-range context dependencies with linear computational complexity by introducing a magnitude-aware mechanism, while the latter reinforces the semantic density of local features via gating mechanisms. Furthermore, addressing the difficulty of separating small objects, a Sparse Parameter Gaussian-Guided Enhancement (SPGT) mechanism is proposed. Based on information theory, SPGT constructs target-specific saliency ground-truth maps for auxiliary supervision during the training phase, significantly improving the model's sensitivity to small objects without increasing inference costs. Extensive experiments on three public datasets, including: VisDrone2019, USOD, and RSOD, demonstrate that GRO-YOLO, with only approximately 6 M parameters, outperforms state-of-the-art methods such as YOLO11 and FFCA-YOLO in both detection accuracy and inference speed, achieving balance between effectiveness and efficiency.
No takes yet. Share an insight, caveat, or question.
Li et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: