小样本图像分类中的类别信息融合网络

Category Information Fusion Network in Few-Shot Image Classification

小样本图像分类任务要求模型仅从少量的图像样本中学到新类别的正确分类方法,是一种特殊的分类任务。然而,以往大多数小样本工作都单独处理来自不同类别的样本,而没有充分利用到不同类别间的信息。本文提出了一种新的类别融合网络(Category-fusion network,CFN),通过融合来自不同类别的样本信息,同时挖掘类别内和类别间的信息。CFN的重要部分是一个融合映射的学习,即如何融合样本中的特征,从而映射出网络参数。其中的一个重要问题是融合映射是否应该随不同的输入样本而改变。本文设计了3个不同的模块:具有固定映射的类无关模块、融合映射仅依赖于期望学习的目标类别的半相关模块和完全相关的模块,其中融合映射完全依赖于输入样本。本文的网络可以通过学习多个类别的样本之间的关系来进行类别概念的学习,并生成融合信息的分类器。实验结果表明,本文网络在广泛应用的MiniImageNet数据集上得到了60.03%的分类精度。

Few-shot image classification is a special task where the model learns to build correct concepts of categories from only a few examples.Due to the frequent occurrence of few-shot scenarios,it has aroused extensive research.However,most previous few-shot models process examples from different categories individually without considering inter-classes information.We propose a novel category-fusion network(CFN)to exploit the intra-and inter-classes information simultaneously by fusing the information of examples from different categories.The key part of CFN is the learning of a fusion map,that is,how to fuse the features in the sample to map out the network parameters.There is an important problem that whether the fusion mapping should change with different input examples.To explore this problem,we design three different modules:(1)the class-irrelevant module with a fixed mapping;(2)the semi-relevant module where the fusion mapping only depends on the target category whose knowledge is expected to be learned;(3)the fullyrelevant module where the fusion mapping totally depends on input examples.Our network can build the concept of a certain category by learning from examples of several categories,and generates a classifier with fused information.The experiments show the effectiveness of our network in fewshot learning,which obtains 60.03% in accuray on the widely used MiniImageNet dataset.