SGDet-Light: Synergistic Global-Local Learning for Efficient Small Object Detection

To address the critical challenge of spatial information degradation in low-resolution small objects and prohibitive computational costs hindering mobile deployment, this work proposes SGDet-Light —a novel lightweight detection framework integrating Cross-layer Global Attention (CGA) and SandGlass Bottleneck Convolution (SGB). First, a hierarchical CGA mechanism is designed to establish inter-layer contextual dependencies, effectively suppressing feature redundancy while mitigating over-fitting across diverse scenarios. Second, an SandGlass bottleneck convolution (SGB) architecture with dual-path differentiable operators — Enhanced SandGlass Convolution (ESConv) for spatial-identity preservation in deep layers and Fused SandGlass Convolution (FSConv) for parameter efficiency in shallow layers to resolves gradient confusion and accelerates training convergence. Finally, we introduce an adaptive spatial feature fusion technique as the model’s multi-head predictor, en Experimental results demonstrate that our method outperforms state-of-the-art approaches on the MS COCO 2017 Val, Pascal VOC 2012 Val, and DOTA datasets, achieving improvements of 1.3% in AP, 1.4% in AP50, and 2.2% in mAP@0.5, while reducing parameters by 5.4% compared to the lightweight MobileNetV3.The framework achieves an optimal accuracy-efficiency trade-off, enabling real-time inference on edge devices.