基于自适应特征比较的少样本学习算法

Few-shot Learning Algorithm Based on Adaptive Feature Comparison

针对少样本学习问题,提出基于自适应特征比较的算法.通过元学习的策略,在训练集中以基于自适应特征比较的方式学习到可用于直接判断查询图像与每一张训练图像类别相似度分数的知识;利用学习到的知识,在测试过程中将所有支撑图像与查询图像经过一遍前向计算,选取与查询图像经特征比较最相似的一幅支撑图像的类别作为查询图像的类别.算法由特征提取和特征比较两个卷积神经网络子模型组成,分别实现图像空间到特征空间的转换和比较两个特征属于同一类别的相似度.两个子模型组合成一个统一的网络通过端到端的方式进行训练,在两个常用少样本学习公开评测集Omniglot和miniImagenet上的实验证明了提出的算法能够提升少样本学习的性能.

In this paper,aiming at the problem of few-shot learning,an algorithm based on adaptive feature comparison is proposed.Through the strategy of meta learning,the knowledge which can be used to directly judge the similarity score between the query image and each training image is learned in the training set based on adaptive feature comparison.By using the knowledge learned,all supporting images and query images are calculated forward once in the test process,and the category of the support image which is most similar to the query image by feature comparison is selected as the query image category.The proposed algorithm consists of feature extraction and feature comparison,which are two sub models of convolution neural network,which can realize the transformation from image space to feature space and compare the similarity of two features belonging to the same category.The two sub-models are combined into a unified network,which is trained in an end-to-end way.Experiments on two commonly used public evaluation datasets of Omniglot and miniimagenet show that the proposed algorithm can improve the performance of few-shot learning.