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

Wu Y.; Yu C.; Chen J.; Liu X.; Wang X. · Frontiers in plant science · 2026

**Background & Methods** This study develops CBD-YOLO, a lightweight deep learning detection algorithm based on YOLOv9t architecture, designed to identify bud-stage Hangzhou white chrysanthemums in complex field environments. The algorithm incorporates a normalized weighted fusion module for adaptive cross-scale feature integration and a lightweight contextual enhancement module to improve detection of small, densely distributed objects with significant background interference. **Key Findings** - CBD-YOLO achieved superior detection performance with 85.85% precision, 78.80% recall, 89.88% mAP@0.5, and an F1-score of 82.08%, demonstrating improvements of 3.29, 1.83, 2.13, and 2.62 percentage points respectively over the baseline YOLOv9t model. - The model maintained computational efficiency with only 1.976 million parameters, 7.60 GFLOPs computational requirements, and a compact model size of 4.70 MB, enabling practical deployment on edge devices. - Embedded deployment on Jetson NX platform achieved 97.89 frames per second with mean inference latency of 10.74 milliseconds following TensorRT FP16 optimization, demonstrating real-time processing capability suitable for automated harvesting systems. **Dosage & Administration** Not reported. **Safety & Adverse Effects** Not reported. **Evidence Quality** This review represents a technical computer vision study rather than a clinical investigation, with significant limitations for medical application. The research focuses on agricultural automation for chrysanthemum harvesting without direct human health implications. Evidence quality is limited by: (1) absence of human subjects or clinical endpoints; (2) narrow application scope to a single plant species and specific phenological stage; (3) lack of comparison with alternative detection methods beyond YOLOv9t baseline; (4) failure-case analysis acknowledged but not comprehensively detailed; (5) field validation conducted in presumably controlled conditions without demonstration across diverse geographic or seasonal contexts. The study provides Level 4 evidence (technical validation study) rather than clinical evidence. While the algorithm demonstrates promising performance metrics for agricultural robotics applications, direct extrapolation to medical/cannabis research contexts is inappropriate without substantial adaptation and validation.

More recent cannabis research