YOLOv8-SGCC: improved YOLOv8-based detection algorithm for Rosa roxburghii

Reliable detection of Rosa roxburghii in orchard environments presents a critical challenge for automated harvesting systems, where conventional vision techniques struggle with dynamic lighting and occlusion. Current approaches dependent on handcrafted features demonstrate three key limitations: suboptimal processing speeds, vulnerability to environmental variations, and inadequate precision for small targets. Our solution, YOLOv8-SGCC, implements four strategic enhancements to the baseline YOLOv8 architecture to address these constraints. The framework first employs SPD-Conv for non-destructive downsampling, maintaining fine-grained channel information critical for detecting diminutive fruits. Subsequent integration of GAM attention blocks optimizes cross-region feature correlation while suppressing noise propagation. For feature reconstruction, we implement CARAFE's adaptive upsampling rather than conventional interpolation, preserving structural details during resolution enhancement. The neck network further incorporates C2f-ME blocks to streamline feature fusion through multi-scale convolution, achieving parameter efficiency without compromising representational capacity.Benchmarking reveals significant improvements across all metrics: a 4% boost in mAP50 (95.4% vs 91.4%), 6.8% fewer parameters (5.5M), and 5.3% lower FLOPs (7.2G) while processing 108 frames/second. The system particularly excels in challenging scenarios involving immature fruits (94.5% mAP50) and occluded clusters (96.6% mAP50 for mature specimens), outperforming both classical (Faster R-CNN) and contemporary (RT-DETR) detectors in speed-accuracy tradeoffs. These advancements establish a new state-of-the-art for agricultural vision systems targeting delicate produce in uncontrolled environments.