基于注意力机制和图卷积的小样本分类网络

深度神经网络在有着大量标注数据的图像识别任务上已经占据了统治地位,但是在只有少量标注数据的数据集上训练一个好的网络仍然是一个据有挑战性的任务.如何从有限的标注数据中学习已经成为了一个有着很多应用场景的热点问题.目前有很多解决小样本分类任务的方法,但是仍然存在识别准确率低的问题,根本原因是在小样本学习中,神经网络只能接收少量有标签的数据,导致神经网络不能获取足够的用来识别的信息.因此,提出了一种基于注意力机制和图卷积网络的小样本分类模型.这个模型不仅能够更好地提取特征,而且能够充分利用提取的特征对目标图像进行分类.通过注意力机制,能够指导神经网络关注更有用的信息,而图卷积使得网络能够利用支撑集中其他类别的信息做出更准确的判断.经过大量的实验,证明了提出的模型在Omniglot数据集和mini-ImageNet数据集上的分类准确率都超过了基于传统神经网络的关系网络.

Relation Network Based on Attention Mechanism and Graph Convolution for Few-Shot Learning

Deep neural networks have dominated image recognition task with large amounts of labeled data. But training a well-performing network on a smaller dataset is still a very challenging task. How to learn from limited labeled data is a key research with excellent scenarios and potential applications. There are many ways to solve few-shot recognition problem, but there is still a problem of low recognition accuracy. The fundamental reason is that in few-shot learning, the traditional neural network can only accept a small amount of labeled data, which makes the network unable to obtain enough information for identification. Therefore, the paper proposes a few-shot classification model based on attention mechanism and graph convolutional neural network, which can not only extract features better, but also make full use of the features to classify the target image. Through the attention mechanism, it can guide the neural network to pay attention to more useful information, and graph convolution enables the network to make more accurate judgments by using the information from other classes of support set. Through many experiments, it is proved that the classification accuracy of the model on the Omniglot dataset and the miniImageNet dataset surpasses the original relational network which based on traditional neural network.