SCG-YOLO: Shift-Context Network With Multidirectional Upsampling and Geometric Quality for Small-Object Detection

Abstract

With the ongoing progress in remote sensing object detection, small-object detection in complex environments still faces difficulties. The representation of small-object features is often limited, and the localization accuracy can be insufficient, which continues to constrain further improvements in overall performance. To address these challenges, we introduce SCG-YOLO, an improved framework based on YOLO11. First, a multidirectional shift channel upsampling (MSCU) module applies a shift channel mix mechanism. By performing directional cyclic shifts and reorganizing features across grouped channels, MSCU recovers high-frequency spatial details lost in downsampling. Second, the C3K2-structured context-guided block (C3K2-CGB) is built by replacing the bottleneck in the backbone’s C3K2 block with a context-guided block (CGB). Using a dual-branch design to capture local details and broader context separately, followed by global refinement, CGB enhances small-object representation in complex backgrounds. Finally, the swish-gated linear unit (SwiGLU)-based localization quality module is integrated into the detection head, where its more stable gradient behavior and more effective feature gating help improve the reliability of bounding box quality assessment for small objects, thereby supporting more accurate localization. On the NWPU-10 and SIMD datasets, SCG-YOLO achieves mAP@0.5 of 85.5% and 82.3%, representing gains of 3.2 and 1.6 percentage points over the baseline.