Authors
Yong Niu, Jun Ma, Jinyu Fan, Junyi Wu, Haiming Lan
Published in
Science progress. Volume 109. Issue 4. Pages 368504261506752. Epub Nov 12, 2026.
Abstract
Plant disease images commonly suffer from pixel confusion, texture blurring and semantic ambiguity caused by cluttered backgrounds, noise interference and diverse lesion morphologies. These intra- and inter-class disturbances break the representation balance between global context and local details, severely restricting classification accuracy, while reliable disease recognition is critical for food security. Conventional recognition models are mostly trained on single-scene datasets, yielding satisfactory performance only under specific environments with poor cross-scene generalization ability. To tackle this challenge, we propose a Cross-level Spatial-Frequency Collaborative Network (CSFCNet), comprising three synergistic modules for feature extraction, fusion and enhancement. First, the Cross-level Multi-scale Feature Extraction (CLMFE) module integrates 1× 1, 3× 3, 5× 5 and 7× 7 convolutions with a coupled spatial-frequency self-attention block, simultaneously capturing shallow details, mid-level textures and deep semantics to generate fine-grained and semantically complete cross-level representations. Second, the Differential Perception Spatial-Frequency Collaborative Fusion (DPSFCF) mechanism amplifies lesion-specific discriminative differences via feature subtraction while suppressing redundant common components. It further transforms spatial features into the frequency domain via Fast Fourier Transform and applies attention-based modulation to the amplitude spectrum, which suppresses low-frequency background interference and emphasizes lesion-related high-frequency cues, and fuses spatial and frequency information through self-attention for complementary promotion. Finally, a Grouped Dual-Channel Coordinate Attention (GDCCA) module is introduced to focus on disease-related regions and suppress background interference. Experiments on apple, maize and tomato disease datasets show that CSFCNet improves accuracy, precision, recall and F1-score by 1.3-2.5 percentage points over the ResNet-50 baseline, with accuracy gains of 1.48-2.37 percentage points across the three datasets. Under the same from-scratch training protocol, it outperforms vision Transformer baselines by 1.6-17.6 percentage points in accuracy, with the largest gains for pure Transformer architectures. It reliably detects tiny lesions in complex scenes, reduces pesticide waste and yield loss from misdiagnosis, and provides technical support for precision plant protection.
PMID:
43020629
Bibliographic data and abstract were imported from PubMed on 13 Nov 2026.
Read full publication at:
Please sign in
to see all details.
Advertisement
Stats
- Recommendations n/a n/a positive of 0 vote(s)
- Views 3
- Comments 0