基于多角度学生子注意力网络的小样本学习

Cooperative few⁃shot learning based on multi⁃angle student sub⁃attention networks

小样本图像分类的准确性取决于神经网络模型对样本数据提取图像表征的能力,为了挖掘出图像更多的细节信息,提出了一种基于多角度学生子注意力网络的小样本分类方法。首先将Conv64所提取的特征作为网络的基础特征,其次构建两个学生分支,使网络从图像位置和通道角度来提取图像的细节信息,最后融入互学习思想,促使两个学生子分支相互监督,相互学习,利用两个学生子分支网络分别对小样本任务进行分类。文中在Mini⁃ImageNet和Tiered⁃ImageNet两个数据集上验证了多角度学生子注意力网络有效性,在Mini⁃ImageNet数据集上,该方法5⁃way 1⁃shot准确率为56.54%,5⁃way 5⁃shot准确率为73.87%。在Tiered⁃ImageNet数据集上,该方法5⁃way 1⁃shot及5⁃way 5⁃shot准确率分别上升到59.62%及77.96%。实验结果表明,相较于只使用单一角度的注意力网络,基于多角度学生子注意力能够更加关注图像的全局信息,显著提高了小样本图像分类的准确性。

The accuracy of few⁃shot image classification depends on the ability of the neural network model to extract image representation from sample data.We propose a few⁃shot learning classification based on multi⁃angle student sub⁃attention networks to mine more detailed information from images.First,Conv64 is used to extract the basic information of the support set and the query set images as the basic features of the network.Second,the network is designed to focus on the position information and channel information of the image through the position and channel sub⁃attention module,and the detail information of the image is extracted from the position and channel.Third,the idea of mutual learning is deployed to promote the mutual supervision and mutual learning of the probability distribution predicted by the two student sub⁃networks.Two student sub⁃networks are used to fulfill few⁃shot tasks.Finally,we conduct experiments on Mini⁃ImageNet and Tiered⁃ImageNet data sets to verify the effectiveness of the multi⁃angle student sub⁃attention network.The results show that the accuracies of the proposed method are 56.54%and 73.87%on Mini⁃ImageNet dataset.The accuracies of 5⁃way 1⁃shot and 5⁃way 5⁃shot are increased to 59.62%and 77.96%,respectively,on Tiered⁃ImageNet dataset.The results demonstrate that,compared with a single attention network,the multi⁃angle student sub attention network pays more attention to the global information of the image,and significantly improve the accuracy of few⁃shot image classification.