4 November 2025 in IEEE Signal Processing Letters
Object detection in aerial remote sensing images is challenging due to small target size, low resolution, and frequent occlusions. To address these issues, this paper proposes SDWS-YOLO, an enhanced model based on YOLOv8. In the backbone, the original C2f modules are replaced with a novel Star operation and Depthwise Separable convolution (SDWS) module, which utilizes three distinct operations to strengthen feature representation. The final C2f module and the last downsampling layer are removed to minimize small object feature loss. For the neck, a Four-FPN structure is introduced, integrating multi-scale features via straightforward channel concatenation. The P5 detection head is eliminated to reduce parameters and computational cost. Furthermore, a composite WFS-IoU loss (WiseIoU-FocalerIoU-ShapeIoU) is adopted to improve localization accuracy. Experiments on the VisDrone2019 and DIOR datasets show that SDWS-YOLO improves mAP50 by 2.9% and 1.3%, respectively, over YOLOv8, while having 1.9 million fewer parameters and lower complexity, achieving a superior performance-efficiency balance.
9 July 2026 in Expert Systems with Applications