基于图网络优化及标签传播的小样本图像分类算法

Few-shot Image Classification Algorithm Based on Graph Network Optimization and Label Propagation

近年来,虽然深度学习技术在图像分类任务中取得了有竞争力的表现,但实际应用中,往往存在缺乏大量训练样本的情况,易于产生过拟合现象。小样本学习技术为此提供了解决方案。由于图神经网络在表示类内和类间样本关系上的优势,已被用于小样本图像分类任务。现有算法是通过几个卷积块获取图像特征作为节点特征输入图网络,为了更好的表示图节点之间的关系,本文算法通过引入卷积块注意力模块(CBAM)来增强目标显著性区域特征,用更具表征力的特征来优化图网络。该方法通过图网络信息传播隐式地增强样本间关系以达到优化节点关系的目的,并采用标签传播机制对未知样本进行分类。大量实验表明,本算法在小样本监督和半监督学习任务上具有优异的表现,验证了算法的有效性。

In recent years,deep learning technology has achieved competitive performance in image classification tasks,but in practical applications,there is often a lack of a large number of training samples,which may result in over-fitting phenomenon.Due to its advantages in representing inter and intra relationships of different samples,of graph neural network has been used for few-shot image classification tasks.The existing algorithms often use a few convolutional blocks to obtain image features as the node feature input graph network.In order to better represent the relationship between the graph nodes,the algorithm in this paper introduces the convolutional block attention module(CBAM)to enhance the saliency of the target regional features,use more representative features to optimize the graph network.This method implicitly enhances the relationship between samples through the spread of graph network information to achieve the purpose of optimizing node relationships.After that,the label propagation mechanism is used to classify unknown samples.Intensive experimental results proved the excellent performance of the proposed algorithm in few-shot supervised and semi-supervised learning tasks,and verified its effectiveness.