摘要
针对目前在小样本图像分类任务上,现有模型存在特征提取不足、提取的特征信息单一等问题,提出一种结合图神经网络的双特征提取的原型网络(Graph-Based Dual-Encoding Prototype Network,GB-ProtoNet)。GB-ProtoNet采用了双特征提取器结构,通过整合2种不同的神经网络,有效地捕获并融合了样本的全局特征信息和局部特征信息。具体而言,该模型结合了残差网络(Residual Network,ResNet)和基于图谱卷积网络(Graph SAmple and aggreGatE,GraphSAGE)的图神经网络。ResNet能够在多层网络结构中有效地传递和保留信息,从而提取样本的全局特征信息。GraphSAGE擅长处理图结构数据,通过采样和聚合邻居节点的信息,提取出样本的局部特征信息。GB-ProtoNet在训练阶段使用标签平滑交叉熵损失函数计算损失并更新模型参数。该模型在miniImageNet和CUB-2002-2011数据集上进行对比实验,与其他经典模型相比,GB-ProtoNet在5-Way 1-Shot和5-Way 5-Shot两种设置上均取得最佳分类精度。
To address the limitations of existing models in capturing comprehensive and diverse features for few-shot image classification tasks,a Graph-Based Dual-Encoding Prototype Network(GB-ProtoNet)is proposed,which employs a dual-feature extractor architecture that effectively captures and fuses global and local feature information by integrating two distinct neural networks.Specifically,a Residual Neural Network(ResNet)is combined with a graph neural network based on Graph SAmple and aggreGatE(GraphSAGE).The ResNet effectively propagates and preserves information across multiple network layers,enabling the extraction of global feature information.The GraphSAGE,on the other hand,shows good performance in processing graph-structured data by sampling and aggregating information from neighboring nodes to extract local feature information.During training,the GB-ProtoNet utilizes label-smoothing cross-entropy loss function to compute the losses and update the model parameters.Extensive experiments conducted on the miniImageNet and CUB-2002-2011 datasets demonstrate that the GB-ProtoNet achieves superior classification accuracy compared to other state-of-the-art models under both 5-Way 1-Shot and 5-Way 5-Shot settings.
作者
齐露露
俞卫琴
QI Lulu;YU Weiqin(School of Mathematics and Statistics,Shanghai University of Engineering Science,Shanghai 201600,China)
出处
《无线电工程》
2024年第7期1668-1675,共8页
Radio Engineering
