.png){kind=icon}
.png){kind=icon}
Return
TransFeatWalk: bidirectional prototype-query adaptation with multi-head feature fusion for few-shot plant disease classification
.png){kind=icon}
PRE
AI
R
J
A
DOI：10.1007/s00500-026-11410-y.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Few-shot learning (FSL) has emerged as an effective approach to tackling data scarcity in real-world applications, such as plant disease classification, where obtaining labeled samples is both costly and labor-intensive. Existing FSL methods often depend on static feature representations and unidirectional adaptation mechanisms. It restricts their capacity to capture task-specific discriminative information in varying environmental conditions. To overcome these constraints, this paper presents TransFeatWalk, a FSL framework that incorporates bidirectional prototype-query adaptation to enhance feature alignment. The proposed model facilitates mutual refinement between support prototypes and query features, which enhances the model’s capacity to generalize to unseen classes. A multi-head feature fusion module has also been added to capture different feature interactions across different representation subspaces. This enhancement improves robustness against both intra-class variability and inter-class similarity. The framework further employs a cosine-similarity–based classifier for effective decision-making in low-data regimes. The overall architecture works in a transductive setting, using query information during inference to improve class representations. Experiments are conducted on the Cashew–Cassava–Maize–Tomato (CCMT) and PlantVillage datasets under standard 5-way 1-shot, 5-shot, and 10-shot settings. The proposed model achieves 43.29%, 52.94%, and 57.10% accuracy on CCMT, and 52.06%, 79.49%, and 87.88% on the PlantVillage dataset. These results indicate that combining bidirectional adaptation with multi-head feature fusion improves feature representation and classification performance in low-data environments. TransFeatWalk performs well as a solution for plant disease classification when only a small amount of data is available.
Keywords:
Few-shot learning
Transductive learning
Plant disease classification
Multi-head feature fusion
Image classification
Journal
IF:
2.5
Papers:
1.0W
Citations:
2.1W
Organization
Cited Papers