Lightweight YOLOv8n marigold corolla detection and picking-point localization method based on ECR progressive channel compression and dual attention mechanisms
Abstract
To address insufficient detection accuracy and picking-point positioning difficulties from dense occlusion of marigold corollas in fields, we propose a lightweight detection and spatial localization method. A Femto Mega depth camera acquires top-view color and depth images. For detection, we adopt YOLOv8n, introduce AIFI (Attention-based Intra-scale Feature Interaction) to replace SPPF for enhancing global context, employ C2fAttn (C2f with Attention) dynamic feature weighting in neck to replace C2f for strengthening local detail on dense overlaps, and propose ECR (Efficient Channel Reduction) progressive compression for lightweighting via attention-compensated channel loss. For localization, we propose a picking-point positioning algorithm using depth and spherical fitting, with centroid offset correction and sphere intersection for 3D coordinates. Experiments show model parameters 1.2M (60% compression) and [email protected] 98.7%; average localization errors 4.5 mm indoors and 7.5 mm outdoors, with system error 2–18 mm. When corolla stem inclination exceeds 30°, the top-view spherical fitting fails, which is the main constraint in unstructured fields. At tilt angles not exceeding 30°, the positioning accuracy met the tolerance requirements of the end-effector of the harvesting manipulator, and the theoretical picking point was consistently located below the actual harvestable position. These results indicate that the proposed method maintains high accuracy while substantially reducing model complexity, thereby providing a feasible solution for intelligent harvesting of marigolds.
Related
- GFLNet: A lightweight real-time detector with knowledge-guided compression in remote sensing images
- LTK-YOLO: Lightweight Keypoint Detection for Robotic Tea Bud Picking
- LeafXNet: A Lightweight Explainable Model With Integrated Attention Mechanism for Fruit Leaf Disease Detection
- Deep learning-based detection of construction waste in complex scenarios with an improved lightweight algorithm
- A temperature-adaptive logit-standardized knowledge distillation for lightweight network intrusion detection
BibTeX
@article{wang2026lightweight,
title = {{Lightweight YOLOv8n marigold corolla detection and picking-point localization method based on ECR progressive channel compression and dual attention mechanisms}},
author = {Yingfan Wang and Baojian Ma and Zhenghao Wu and Yun Ge and Bangbang Chen},
journal = {Frontiers in Plant Science},
year = {2026},
month = nov,
doi = {10.3389/fpls.2026.2005832},
url = {https://www.frontiersin.org/articles/10.3389/fpls.2026.2005832},
note = {CC BY 4.0},
}
RIS
TY - JOUR
TI - Lightweight YOLOv8n marigold corolla detection and picking-point localization method based on ECR progressive channel compression and dual attention mechanisms
AU - Yingfan Wang
AU - Baojian Ma
AU - Zhenghao Wu
AU - Yun Ge
AU - Bangbang Chen
PY - 2026
DA - 2026/11/05
JO - Frontiers in Plant Science
DO - 10.3389/fpls.2026.2005832
UR - https://www.frontiersin.org/articles/10.3389/fpls.2026.2005832
AB - To address insufficient detection accuracy and picking-point positioning difficulties from dense occlusion of marigold corollas in fields, we propose a lightweight detection and spatial localization method. A Femto Mega depth camera acquires top-view color and depth images. For detection, we adopt YOLOv8n, introduce AIFI (Attention-based Intra-scale Feature Interaction) to replace SPPF for enhancing global context, employ C2fAttn (C2f with Attention) dynamic feature weighting in neck to replace C2f for strengthening local detail on dense overlaps, and propose ECR (Efficient Channel Reduction) progressive compression for lightweighting via attention-compensated channel loss. For localization, we propose a picking-point positioning algorithm using depth and spherical fitting, with centroid offset correction and sphere intersection for 3D coordinates. Experiments show model parameters 1.2M (60% compression) and [email protected] 98.7%; average localization errors 4.5 mm indoors and 7.5 mm outdoors, with system error 2–18 mm. When corolla stem inclination exceeds 30°, the top-view spherical fitting fails, which is the main constraint in unstructured fields. At tilt angles not exceeding 30°, the positioning accuracy met the tolerance requirements of the end-effector of the harvesting manipulator, and the theoretical picking point was consistently located below the actual harvestable position. These results indicate that the proposed method maintains high accuracy while substantially reducing model complexity, thereby providing a feasible solution for intelligent harvesting of marigolds.
SN - 1664-462X
ER -