Xinyi Lin
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Xinyi Lin has a total of 7 co-authors. Here are the top co-authors by number of shared publications:
- Fengjun Zhou
School of Computer Science, Shandong Xiehe University, Jinan, China
1 shared publication - Guodong Wang
Shandong Xiehe University, No. 6277, Jiqing Road, Licheng District, Jinan City, Shandong Province, China
1 shared publication - Qianqian Li
Sinopec Sales Co., Ltd., Shandong Binzhou Petroleum Branch, Binzhou 256600, P. R. China
1 shared publication - Qun Wang
Shandong Xiehe University, Jinan, P. R. China
1 shared publication - Weiguo Lai
Shandong Xiehe University, Jinan, P. R. China
1 shared publication - Xueqiang Gao
Shandong Xiehe University, No. 6277, Jiqing Road, Licheng District, Jinan City, Shandong Province, China
1 shared publication
Latest Publications
- 23 July 2025
- Published by World Scientific Pub Co Pte Ltd in International Journal of Pattern Recognition and Artificial Intelligence
A Transformer-Based Method for Few-Shot Image Recognition in Federated Learning Environments
Few-shot image recognition in federated learning (FL) environments presents unique challenges due to data heterogeneity, limited local samples, and strict privacy constraints. To address these issues, we propose a novel Transformer-based approach that enables effective visual recognition with minimal labeled data on distributed clients. The proposed model introduces an adaptive multi-scale window attention mechanism to enhance feature extraction under sparse data conditions. Furthermore, a collaborative attention strategy is developed to enable privacy-preserving semantic knowledge exchange across clients. To alleviate the class imbalance and overfitting problems common in few-shot settings, we design an improved Focal Loss function guided by class-wise entropy. Additionally, a progressive knowledge distillation framework is incorporated to maintain model consistency across training rounds. Extensive experiments conducted on multiple few-shot image benchmarks under federated settings demonstrate that our method significantly outperforms existing baselines in terms of accuracy, communication efficiency, and generalization.
- 23 July 2025
- Published by World Scientific Pub Co Pte Ltd in International Journal of Pattern Recognition and Artificial Intelligence
A Transformer-Based Method for Few-Shot Image Recognition in Federated Learning Environments
Few-shot image recognition in federated learning (FL) environments presents unique challenges due to data heterogeneity, limited local samples, and strict privacy constraints. To address these issues, we propose a novel Transformer-based approach that enables effective visual recognition with minimal labeled data on distributed clients. The proposed model introduces an adaptive multi-scale window attention mechanism to enhance feature extraction under sparse data conditions. Furthermore, a collaborative attention strategy is developed to enable privacy-preserving semantic knowledge exchange across clients. To alleviate the class imbalance and overfitting problems common in few-shot settings, we design an improved Focal Loss function guided by class-wise entropy. Additionally, a progressive knowledge distillation framework is incorporated to maintain model consistency across training rounds. Extensive experiments conducted on multiple few-shot image benchmarks under federated settings demonstrate that our method significantly outperforms existing baselines in terms of accuracy, communication efficiency, and generalization.