Discriminative learning of imaginary data for few-shot classification

초록

Humans can quickly learn new visual categories, because they can easily visualize or imagine what we need to recognize in an image with a complex background, and how it differs from other images. Incorporating this ability to hallucinate discriminative instances of novel classes might help machine vision systems perform better few-shot classification. In this paper, we propose a Discriminative Hallucination Learning Network based on attentional mechanism, and unify the GNN-based classifier with a "hallucinator" for few-shot classification. Firstly, we use a pre-trained saliency network to hallucinate the foreground image. Then, the hallucination feature network (FNet) and zoom network (ZNet) are designed to extract more fine-grained local images adaptively with intra-cluster similarity and the inter-cluster dissimilarity. The embedding network (Enet) initialize the node representation of the graph structure from a jointly trained convolutional neural network. Finally, the proposed method is evaluated extensively on three challenging ZSL benchmark datasets. It significantly outperforms state-of-the-art methods in both supervised and semi-supervised few-shot image classification experiments. (c) 2021 Elsevier B.V. All rights reserved.

키워드