Few-shot Self-optimization Learning Based on Deep Metric

Abstract

Due to the lack of labeled images and the effects of noisy pseudo-labels make few-shot learning (FSL) extremely challenging. For this problem, we propose a deep metric self-optimization (DMSO) method based on deep neural networks. Firstly, we construct a siamese network classifier (SNC) that was trained through labeled images. Then, the classifier is used to predict pseudo-labels for unlabeled images. In order to reduce the influence of noise, we create a confidence selection module (CSM) and use it to select pseudo-labeled images with higher confidence. Then, adding pseudo-labeled images to the training set. Finally, we use labeled images and pseudo-label images to jointly retrain the classifier. By the above means, we achieve the purpose of data augmentation and model self-optimization. In this paper, we focus on the one-shot classification task and achieve the state-of-the-art results which exceed those of other few-shot learning models. We evaluate DMSO method on two FSL benchmarks including miniImageNet and tieredImageNet.