Metric-Based Learning for Nearest-Neighbor Few-Shot Image Classification

Abstract

Few-shot learning task, which aims to recognize a new class with insufficient data, is an inevitable issue to be solved in image classification. Among recent work, Metalearning is commonly used to Figure out few-shot learning task. Here we tackle a recent method that uses the nearest-neighbor algorithm when recognizing few-shot images and to this end, propose a metric-based approach for nearest-neighbor few-shot classification. We train a convolutional neural network with miniImageNet applying three types of loss, triplet loss, crossentropy loss, and combination of triplet loss and cross-entropy loss. In evaluation, three configurations exist according to feature transformation technique which are unnormalized features, L2-normalized features, and centered L2-norma1ized features. For 1-shot 5-way task, the triplet loss model attains the uppermost accuracy among all three configurations and for 5-shot 5-way task, the identical model reaches the foremost accuracy in unnormalized features configuration.