LT-YO-AN: lightweight YOLOv12 with asymmetric neck for apple fruitlet detection in complex orchard environments
- 1College of Communication and Information Engineering, Xi’an University of Science and Technology
- 2Xi’an Key Laboratory of Network Convergence Communication
- 3School of Automation Science and Engineering, Faculty of Electronic and Information Engineering, Xi’an Jiaotong University
Abstract
Accurate detection of apple fruitlets in complex orchard environments is a prerequisite for precision fruit thinning but remains challenging due to small target sizes, variable illumination, and dense foliage. To address these issues, a lightweight YOLOv12 architecture, termed LT-YO-AN, was proposed. The backbone integrates RepDown re-parameterizable downsampling and D3C2f aggregation modules, enabling hierarchical feature extraction with reduced spatial redundancy and channel-level compression. An asymmetric neck is introduced to explicitly enhance the high-resolution P2 layer for small-target sensitivity, while deliberately weakening the P5 layer via channel reduction (256 to 128) to alleviate redundancy. This design is complemented by a four-scale detection head (P2–P5) and a width scaling strategy (factor = 0.25), which together allocate computational resources preferentially to shallow features, yielding a compact model optimized for apple fruitlet detection. Evaluations on a real-world orchard dataset demonstrate that LT-YO-AN achieves an mAP50–95 of 0.675 and a recall of 0.860, outperforming benchmark YOLO variants. Although the computational complexity increases to 11.5 GFLOPs, the model maintains a compact size of 3.5 MB with 1.49 M parameters, making it suitable for edge deployment. This study provides a viable technological solution for early-stage fruit detection and automated orchard management.
Related
- Cross-level attention and lightweight feature extraction for improved tomato ripeness detection in complex greenhouse environments
- LTK-YOLO: Lightweight Keypoint Detection for Robotic Tea Bud Picking
- LeafXNet: A Lightweight Explainable Model With Integrated Attention Mechanism for Fruit Leaf Disease Detection
- Chili-Seg: a high-precision detection and instance segmentation model for chili flowers in complex environments
- Deep learning-based detection of construction waste in complex scenarios with an improved lightweight algorithm
BibTeX
@article{wang2026ltyoan,
title = {{LT-YO-AN: lightweight YOLOv12 with asymmetric neck for apple fruitlet detection in complex orchard environments}},
author = {Dandan Wang and Bo Wang},
journal = {Frontiers in Plant Science},
year = {2026},
month = oct,
doi = {10.3389/fpls.2026.1958019},
url = {https://www.frontiersin.org/articles/10.3389/fpls.2026.1958019/full},
note = {CC BY 4.0},
}
RIS
TY - JOUR
TI - LT-YO-AN: lightweight YOLOv12 with asymmetric neck for apple fruitlet detection in complex orchard environments
AU - Dandan Wang
AU - Bo Wang
PY - 2026
DA - 2026/10/14
JO - Frontiers in Plant Science
DO - 10.3389/fpls.2026.1958019
UR - https://www.frontiersin.org/articles/10.3389/fpls.2026.1958019/full
AB - Accurate detection of apple fruitlets in complex orchard environments is a prerequisite for precision fruit thinning but remains challenging due to small target sizes, variable illumination, and dense foliage. To address these issues, a lightweight YOLOv12 architecture, termed LT-YO-AN, was proposed. The backbone integrates RepDown re-parameterizable downsampling and D3C2f aggregation modules, enabling hierarchical feature extraction with reduced spatial redundancy and channel-level compression. An asymmetric neck is introduced to explicitly enhance the high-resolution P2 layer for small-target sensitivity, while deliberately weakening the P5 layer via channel reduction (256 to 128) to alleviate redundancy. This design is complemented by a four-scale detection head (P2–P5) and a width scaling strategy (factor = 0.25), which together allocate computational resources preferentially to shallow features, yielding a compact model optimized for apple fruitlet detection. Evaluations on a real-world orchard dataset demonstrate that LT-YO-AN achieves an mAP50–95 of 0.675 and a recall of 0.860, outperforming benchmark YOLO variants. Although the computational complexity increases to 11.5 GFLOPs, the model maintains a compact size of 3.5 MB with 1.49 M parameters, making it suitable for edge deployment. This study provides a viable technological solution for early-stage fruit detection and automated orchard management.
SN - 1664-462X
ER -