A lightweight intelligent leaf disease recognition method based on asymmetric factorized attention: A case study on apple leaves
Physiological and Molecular Plant Pathology · 2027-01 · 147 · 103541
DOI: 10.1016/j.pmpp.2026.103541 ↗Kurz gefasst
Fachartikel in der Zeitschrift Physiological and Molecular Plant Pathology, 2027-01. Verfasst von Janbo Yue und 2 weiteren Autor:innen. Die Arbeit nennt 30 Literaturangaben.
Abstract (Originalsprache)
Against the backdrop of the rapid development of smart agriculture, real-time automated identification of leaf diseases has become a key technological support for ensuring global food security and achieving precision plant protection. However, traditional deep learning models struggle to balance computational complexity with fine-grained feature representation under resource-constrained edge environments. To bridge this gap, a lightweight method is proposed for intelligent leaf disease identification based on Asymmetric Factorized Attention (FAB), which mitigates Multi-Head Self-Attention (MHSA) computational redundancy through a spatial-channel decoupling strategy. Unlike conventional symmetric attention variants where query, key, and value share unified dimensions, FAB factorizes the query channel dimension via compression to the head count, downscales the key spatial size via adaptive pooling, and preserves the full-dimensional Value carrier, capturing subtle lesion variations with lower matrix-multiplication overhead. The FAB module is integrated into the Transformer encoder of MobileViTv3 as a contextual sub-layer to capture global dependencies, while a parallel-orthogonal information fusion block, HABlock, is designed within the YOLOv12 backbone to process spatial and channel attention dimensions simultaneously, thereby enhancing multi-scale lesion localization. Validation on an apple leaf disease dataset demonstrates that the proposed classification network achieves a 99.13% Top-1 accuracy with an extremely small footprint of 0.69M parameters and 245M FLOPs. Concurrently, the enhanced detection framework delivers a mAP at 50 of 97.50% while achieving an inference speed of 63.3 frames per second (FPS) under the test workstation environment. This asymmetric factorization mechanism effectively breaks the performance bottleneck on edge terminals, offering an efficient and technically sound solution for crop disease monitoring.
Zitieren
APA
Yue, J., Sun, M., & Liu, Z. (2027). A lightweight intelligent leaf disease recognition method based on asymmetric factorized attention: A case study on apple leaves. Physiological and Molecular Plant Pathology, 147, 103541. https://doi.org/10.1016/j.pmpp.2026.103541
Vancouver
Yue J, Sun M, Liu Z. A lightweight intelligent leaf disease recognition method based on asymmetric factorized attention: A case study on apple leaves. Physiological and Molecular Plant Pathology. 2027;147:103541. doi:10.1016/j.pmpp.2026.103541
BibTeX
@article{Yue2027,
author = {Yue, Janbo and Sun, Mingyu and Liu, Zejin},
title = {A lightweight intelligent leaf disease recognition method based on asymmetric factorized attention: A case study on apple leaves},
journal = {Physiological and Molecular Plant Pathology},
year = {2027},
volume = {147},
pages = {103541},
doi = {10.1016/j.pmpp.2026.103541}
}
EUVIMED