Flora-YOLO: A lightweight framework for real-time rose grading on edge devices via structural optimization and...

Precise flower grading in modern agriculture is critical for determining market value but remains hindered by inefficient manual labor and the computational constraints of edge devices. To address these challenges, this paper introduces Flora-YOLO, a lightweight and robust detection framework built upon the YOLO11n architecture. To overcome the inherent bottlenecks of agricultural vision, the framework incorporates four targeted optimizations, explicitly distinguishing novel components from adapted architectures: (1) an adapted C3k2-based Sandglass-Gated Bottleneck (C3k2_SGB) module to extract and amplify minute floral textures; (2) an integrated Adaptive Downsampling (ADown) module to preserve critical high-frequency spatial cues under severe foliage occlusion; (3) a newly proposed CSP-based Variance-Gated Attention (C2VGA) hybrid attention mechanism to amplify the weak semantic signals of early-stage buds while suppressing background noise; and (4) a novel Focused Shape-Intersection over Union (FSIoU) loss function to provide shape-aware gradient guidance for the precise bounding-box regression of irregular, non-rigid floral boundaries. Evaluated via 5-fold cross-validation on a custom dataset, the model achieves a stable 86.7% mAP@0.5 and 50.4% mAP@0.5:0.95. This represents a 3.4% accuracy improvement over the baseline, achieved alongside a 28.6% reduction in computational load, requiring only 1.73 M parameters and 4.5 GFLOPs. Furthermore, out-of-domain testing and evaluation on public agricultural datasets confirm the model's cross-species generalization. Finally, hardware deployment on both Android terminals and an NPU-accelerated NXP i.MX 8M Plus platform yields a real-time inference speed of 11.05 FPS with a 90.5 ms latency. These results establish Flora-YOLO as a highly efficient, accurate, and scalable solution for automated, edge-based sorting in precision floriculture.