LEA: A Lightweight Local Enhanced Attention Module with Local-Competitive Spatial Recalibration for CNN-Based Agricultural Image Classification
Year: 2026
Type: article
Abstract: Local Enhanced Attention (LEA) adds neighborhood aggregation to a learned spatial gate for CNN-based agricultural image classification. The module combines strip filtering, channel-wise normalization and sigmoid gating, followed by aggregation without additional learned parameters. In a seven-method PlantSeg comparison, LEA recorded the highest mean lesion CAM energy (41.612%) and tied with SE in mean pointing accuracy (52.054%). Its PlantVillage mean accuracy was 99.902%, compared with 99.883% for baseline. The measured ResNet-50 configuration added 0.946% parameters, with higher latency and memory use. Classification, binary CAM overlap and corruption results varied across settings. These results support the use of LEA in agricultural classification workflows that inspect disease-related spatial cues.
Source: Applied Sciences
DOI: 10.3390/app162111085
Authors: Kepeng Luo, Zhanyi Lian, Xin Han, Yunfeng Wang
Institution: Xiamen University
Language: English
Topic: Smart Agriculture and AI
Subfield: Plant Science
Domain: Life Sciences
Open Access status: gold
Estimated APC paid: $2,896
OpenAlex ID: W7225658748
FWCI:
Cites:
Cited by:
Related to: