FAnet: A lightweight end-to-end framework for Fructus Aurantii identification and quality grading
- 1Institute of Agricultural Equipment, Xinjiang Uygur Autonomous Region Academy of Agricultural Sciences, Urumqi, China
- 2Xinjiang Key Laboratory of Intelligent Control Technology for Facility Agriculture, Urumqi, China
- 3College of Engineering in Jiangxi Agricultural University, Jiangxi Provincial Key Laboratory of Modern Agricultural Equipment, Nanchang, China
Abstract
The quality grade of Fructus Aurantii(FA) is closely related to its clinical efficacy and medication safety; therefore, a standardized grading method is needed. However, FA samples show considerable morphological variation, and conventional grading depends largely on manual visual assessment, which is subjective and inefficient. This study proposes FAnet, a lightweight instance segmentation framework based on YOLOv8n-seg, for estimating peel thickness and outer diameter during FA quality grading. FAnet integrates three task-adapted components: C2Star, which combines StarNet and C2f to improve multi-scale feature extraction; the Simple Parameter-Free Attention Module (SimAM), which enhances spatial-channel feature weighting without introducing additional trainable parameters; and a multi-scale dynamic fitting module (MSDFM), which maps segmentation-derived areas to physical dimensions for grading. Using a self-built dataset containing 1,093 original images and 5,035 augmented training images, FAnet achieved 96.12% precision, 98.77% recall, 98.60% mAP@50, and 77.08% mAP@95, with 2.97 M parameters and 7.9 GFLOPs under the present experimental setting. The results indicate that FAnet can support automated FA segmentation and preliminary grading under controlled imaging conditions. Broader validation across larger and more diverse sample populations is still required before deployment-oriented application.
Related
- Dissection of quality variation and hierarchical quality evaluation of Evodiae Fructus under multidimensional factors
- DTL-YOLO: a lightweight model for real-time tobacco leaf damage detection and mechanical harvesting quality assessment
- TEFAR: a lightweight, configurable framework for semi-automatic artefact component classification in EEG and TMS-EEG
- Alfalfa-Yolo: a boundary-aware lightweight framework for real-time weed detection in alfalfa fields
- HECA-MobileNet: an explainable lightweight attention-based deep learning framework for custard apple disease detection
BibTeX
@article{wu2026fanet,
title = {{FAnet: A lightweight end-to-end framework for Fructus Aurantii identification and quality grading}},
author = {LeTian Wu and KeKe Liao and BingWei Song and ShenBo Guo and XingDong Gao and HuiFeng Shi},
journal = {PLOS One},
year = {2026},
month = sep,
doi = {10.1371/journal.pone.0357781},
url = {https://doi.org/10.1371/journal.pone.0357781},
note = {CC BY 4.0},
}
RIS
TY - JOUR
TI - FAnet: A lightweight end-to-end framework for Fructus Aurantii identification and quality grading
AU - LeTian Wu
AU - KeKe Liao
AU - BingWei Song
AU - ShenBo Guo
AU - XingDong Gao
AU - HuiFeng Shi
PY - 2026
DA - 2026/09/28
JO - PLOS One
DO - 10.1371/journal.pone.0357781
UR - https://doi.org/10.1371/journal.pone.0357781
AB - The quality grade of Fructus Aurantii(FA) is closely related to its clinical efficacy and medication safety; therefore, a standardized grading method is needed. However, FA samples show considerable morphological variation, and conventional grading depends largely on manual visual assessment, which is subjective and inefficient. This study proposes FAnet, a lightweight instance segmentation framework based on YOLOv8n-seg, for estimating peel thickness and outer diameter during FA quality grading. FAnet integrates three task-adapted components: C2Star, which combines StarNet and C2f to improve multi-scale feature extraction; the Simple Parameter-Free Attention Module (SimAM), which enhances spatial-channel feature weighting without introducing additional trainable parameters; and a multi-scale dynamic fitting module (MSDFM), which maps segmentation-derived areas to physical dimensions for grading. Using a self-built dataset containing 1,093 original images and 5,035 augmented training images, FAnet achieved 96.12% precision, 98.77% recall, 98.60% mAP@50, and 77.08% mAP@95, with 2.97 M parameters and 7.9 GFLOPs under the present experimental setting. The results indicate that FAnet can support automated FA segmentation and preliminary grading under controlled imaging conditions. Broader validation across larger and more diverse sample populations is still required before deployment-oriented application.
SN - 1932-6203
ER -