基于Swin Transformer和双聚焦相似度的图卷积标签传播网络小样本分类

Few-Shot Classification with Graph Convolutional Label Propagation Network Based on Swin Transformer and Dual-Focus Similarity

摘要: 小样本学习旨在通过少量标注样本高效识别新类别，然而，现有方法在处理数据稀缺性、关系建模精准性以及图结构泛化能力等问题时仍存在局限性。为此，文章提出了一种基于Swin Transformer和双聚焦相似度的图卷积标签传播网络(Swin Transformer-Based Graph Convolutional Label Propagation Network, ST-GCLPN)小样本分类方法。首先，利用Swin Transformer提取输入样本的全局与局部特征，缓解数据稀缺导致的特征不足问题。其次，通过双聚焦相似度综合全局和局部特征关系，构建精准的图结构，并输入图卷积网络(GCN)以捕获局部邻域特征。随后，GCN的输出作为输入传递到标签传播算法(LPA)中，通过全局图结构高效传播标签信息，增强分类泛化能力。最后，计算查询集的预测标签与真实标签之间的交叉熵损失，优化模型参数。实验结果表明，在miniImageNet数据集的5-way 1-shot和5-way 5-shot任务中，所提方法的分类准确率均有提高，显著优于当前主流方法。该方法有效解决了小样本学习中特征不足、关系建模不精准及泛化能力有限等问题，为小样本分类任务提供了一种高效而鲁棒的解决方案。

Abstract: Few-shot learning aims to efficiently recognize new categories with a limited number of annotated samples. However, existing methods still face limitations in addressing issues such as data scarcity, the precision of relational modeling, and the generalization capability of graph structures. To this end, this paper proposes a few-shot classification method based on a Graph Convolutional Label Propagation Network (GCLPN) that incorporates Swin Transformer and dual-focus similarity. Firstly, the Swin Transformer is utilized to extract both global and local features from input samples, alleviating the issue of insufficient features caused by data scarcity. Secondly, the dual-focus similarity integrates global and local feature relationships to construct an accurate graph structure, which is then fed into a Graph Convolutional Network (GCN) to capture local neighborhood features. Subsequently, the output of the GCN is passed as input to a Label Propagation Algorithm (LPA), which efficiently propagates label information through the global graph structure, enhancing the classification generalization capability. Finally, the model parameters are optimized by calculating the cross-entropy loss between the predicted labels and the true labels of the query set. Experimental results demonstrate that the proposed method significantly outperforms current mainstream methods, with improved classification accuracy on the miniImageNet dataset for both 5-way 1-shot and 5-way 5-shot tasks. This method effectively addresses the challenges of feature insufficiency, imprecise relational modeling, and limited generalization capability in few-shot learning, providing an efficient and robust solution for few-shot classification tasks.