CBD-YOLO: a lightweight small-object detection algorithm for bud-stage Hangzhou white chrysanthemum in complex field environments.

Hangzhou white chrysanthemum is widely cultivated, and its bud stage has the highest commercial value. However, the optimal harvesting window is short, while manual harvesting is constrained by labor shortages and high labor costs. Accurate detection of bud-stage chrysanthemums is therefore essential for automated selective harvesting. This study proposes CBD-YOLO, a lightweight detector based on YOLOv9t, to address the small object size, dense distribution, and background interference of bud-stage chrysanthemums in field images. CBD-YOLO incorporates a normalized weighted fusion module at a high-level cross-scale fusion node to adaptively regulate the contributions of different feature branches, together with a lightweight contextual enhancement module to strengthen local feature representation in shallow high-resolution features. In experiments, CBD-YOLO achieved 85.85% precision, 78.80% recall, 89.88% mAP@0.5, and an F1-score of 82.08%, representing improvements of 3.29, 1.83, 2.13, and 2.62 percentage points over YOLOv9t, respectively. The model contained only 1.976 M parameters, required 7.60 GFLOPs, and had a model size of 4.70 MB. Embedded deployment validation on a Jetson NX platform showed that CBD-YOLO achieved 97.89 FPS with a mean model inference latency of 10.74 ms after TensorRT FP16 optimization. Ablation experiments, heatmap visualizations, and failure-case analysis further demonstrated the effectiveness and limitations of the proposed design. Overall, CBD-YOLO improves bud-stage chrysanthemum detection while maintaining a compact and deployment-friendly structure, providing a feasible technical basis for intelligent detection and automated selective harvesting under complex field conditions.