FDUNet-LRSA: efficient frequency-adaptive segmentation for automated crop disease inspection
6 affiliations
- 1College of Smart Agriculture (College of Artificial Intelligence), Nanjing Agricultural University
- 2School of Engineering and Physical Sciences, University of Lincoln
- 3Lincoln Institute for Agri-food Technology, University of Lincoln
- 4Guangxi Artificial Intelligence Academic and Guilin University of Electronic Technology Nanning Research Institute
- 5School of Electrical Engineering and Automation, Jiangsu Normal University
- 6Key Laboratory of Sweetpotato Biology and Genetic Breeding, Ministry of Agriculture and Rural Affairs, Xuzhou Institute of Agricultural Sciences in Jiangsu Xuhuai District
Abstract
The transition toward industrialized and automated agriculture requires reliable machine vision for high-throughput crop health monitoring and field decision support. Automated sweetpotato disease inspection, however, remains challenging because field images exhibit substantial variations in leaf scale, appearance, illumination, occlusion, and background complexity. Conventional U-shaped networks rely primarily on fixed convolutional kernels and local feature aggregation, whereas Transformer-based alternatives often introduce considerable computational overhead. This article proposes FDUNet-LRSA, a frequency-adaptive segmentation network that integrates frequency-domain dynamic convolution, low-resolution self-attention, and boundary-aware refinement. Frequency-domain dynamic convolution reconstructs input-dependent spatial kernels to improve feature adaptation under changing imaging conditions. Low-resolution self-attention captures long-range contextual dependencies in a fixed compact feature space, while boundary-aware refinement preserves irregular leaf contours during mask reconstruction. Under the reported evaluation protocol, FDUNet-LRSA achieved a mean test foreground IoU of 0.94 and a Dice coefficient of 0.97, with 3.72 GFLOPs. Additional experiments indicate that task-oriented moderate augmentation improves generalization compared with stronger augmentation. The proposed network provides an efficient visual perception approach for automated crop inspection systems.
Related
- Multistage AI Crop Inspection and Segmentation Framework (MACISF)
- Automatic bridge crack detections: A lightweight segmentation model for UAV inspection
- A Deep Learning Approach for Automated Re-Inspection of False Calls in Mixed-Production Surface Mount Lines Under Extreme Class Imbalance
- Fully automated diaphragm segmentation on routine 2D-cine and 3D thoracic MRI: multireader validation
- Organotchi: an open-source tool for the automated segmentation, morphological characterization, and longitudinal tracking of organoids
BibTeX
@article{han2026fdunetlrsa,
title = {{FDUNet-LRSA: efficient frequency-adaptive segmentation for automated crop disease inspection}},
author = {Ru Han and Lei Shu and Grzegorz Cielniak and Yan Zhu and Jucheng Yang and Fan Yang and Weihan Song and Qian Zhang and Kai Huang},
journal = {Frontiers in Plant Science},
year = {2026},
month = oct,
doi = {10.3389/fpls.2026.1963106},
url = {https://www.frontiersin.org/articles/10.3389/fpls.2026.1963106/full},
note = {CC BY 4.0},
}
RIS
TY - JOUR
TI - FDUNet-LRSA: efficient frequency-adaptive segmentation for automated crop disease inspection
AU - Ru Han
AU - Lei Shu
AU - Grzegorz Cielniak
AU - Yan Zhu
AU - Jucheng Yang
AU - Fan Yang
AU - Weihan Song
AU - Qian Zhang
AU - Kai Huang
PY - 2026
DA - 2026/10/16
JO - Frontiers in Plant Science
DO - 10.3389/fpls.2026.1963106
UR - https://www.frontiersin.org/articles/10.3389/fpls.2026.1963106/full
AB - The transition toward industrialized and automated agriculture requires reliable machine vision for high-throughput crop health monitoring and field decision support. Automated sweetpotato disease inspection, however, remains challenging because field images exhibit substantial variations in leaf scale, appearance, illumination, occlusion, and background complexity. Conventional U-shaped networks rely primarily on fixed convolutional kernels and local feature aggregation, whereas Transformer-based alternatives often introduce considerable computational overhead. This article proposes FDUNet-LRSA, a frequency-adaptive segmentation network that integrates frequency-domain dynamic convolution, low-resolution self-attention, and boundary-aware refinement. Frequency-domain dynamic convolution reconstructs input-dependent spatial kernels to improve feature adaptation under changing imaging conditions. Low-resolution self-attention captures long-range contextual dependencies in a fixed compact feature space, while boundary-aware refinement preserves irregular leaf contours during mask reconstruction. Under the reported evaluation protocol, FDUNet-LRSA achieved a mean test foreground IoU of 0.94 and a Dice coefficient of 0.97, with 3.72 GFLOPs. Additional experiments indicate that task-oriented moderate augmentation improves generalization compared with stronger augmentation. The proposed network provides an efficient visual perception approach for automated crop inspection systems.
SN - 1664-462X
ER -