基于跨域对抗学习的零样本分类

Cross-Domain Adversarial Learning for Zero-Shot Classification

零样本学习旨在识别具有少量、甚至没有训练样本的未见类,这些类与可见类遵循不同的数据分布.最近,随着深度神经网络在跨模态生成方面的成功,使用合成的样本对未见数据进行分类取得了巨大突破.现有方法通过共享生成器和解码器,联合传统生成对抗网络和变分自编码器来实现样本的合成.然而,由于这2种生成网络产生的数据分布不同,联合模型合成的数据遵循复杂的多域分布.针对这个问题,提出跨域对抗生成网络(CrossD-AGN),将传统生成对抗网络和变分自编码器有机结合起来,基于类级语义信息为未见类合成样本,从而实现零样本分类.提出跨域对抗学习机制,引入2个对称的跨域判别器,通过判断合成样本属于生成器域分布还是解码器域分布,促使联合模型中的生成器解码器不断优化,提高样本合成能力.在多个真实数据集上进行了广泛的实验,结果表明了所提出方法在零样本学习上的有效性和优越性.

Zero-shot learning(ZSL)aims to recognize novel categories,which have few or even no sample for training and follow a different distribution from seen classes.With the recent advances of deep neural networks on cross-modal generation,encouraging breakthroughs have been achieved on classifying unseen categories with their synthetic samples.Extant methods synthesize unseen samples with the combination of generative adversarial nets(GANs)and variational auto-encoder(VAE)by sharing the generator and the decoder.However,due to the different data distributions produced by these two kinds of generative models,fake samples synthesized by the joint model follow the complex multi-domain distribution instead of satisfying a single model distribution.To address this issue,in this paper we propose a cross-domain adversarial generative network(CrossD-AGN)to integrate the traditional GANs and VAE into a unified framework,which is able to generate unseen samples based on the class-level semantics for zero-shot classification.We propose two symmetric cross-domain discriminators along with the cross-domain adversarial learning mechanism to learn to determine whether a synthetic sample is from the generator-domain or the decoder-domain distribution,so as to drive the generator decoder of the joint model to improve its capacity of synthesizing fake samples.Extensive experimental results over several real-world datasets demonstrate the effectiveness and superiority of the proposed model on zero-shot visual classification.