基于间接域适应特征生成的直推式零样本学习方法  

Feature Generation Approach with Indirect Domain Adaptation for Transductive Zero-shot Learning

作　　者：黄晟[1,2] 杨万里[1] 张译[1] 张小洪 杨丹[1,3] HUANG Sheng;YANG Wan-Li;ZHANG Yi;ZHANG Xiao-Hong;YANG Dan(School of Big Data&Software Engineering,Chongqing University,Chongqing 401331,China;Key Laboratory of Dependable Service Computing in Cyber Physical Society(Chongqing University),Ministry of Education,Chongqing 400044,China;School of Information Science and Technology,Southwest Jiaotong University,Chengdu 611756,China)

机构地区：[1]重庆大学大数据与软件学院,重庆401331 [2]信息物理社会可信服务计算教育部重点实验室(重庆大学),重庆400044 [3]西南交通大学信息科学与技术学院,四川成都611756

出　　处：《软件学报》2022年第11期4268-4284,共17页Journal of Software

基　　金：国家重点研发计划(2018YFB2101200);国家自然科学基金(61772093,61602068);中央高校基本科研业务费专项资金(2019CDYGYB014)。

摘　　要：近年来,零样本学习备受机器学习和计算机视觉领域的关注.传统的归纳式零样本学习方法通过建立语义与视觉之间的映射关系,实现类别之间的知识迁移.这类方法存在着可见类和未见类之间的映射域漂移(projection domain shift)问题,直推式零样本学习方法通过在训练阶段引入无标定的未见类数据进行域适应,能够有效地缓解上述问题并提升零样本学习精度.然而,通过实验分析发现,这种直接在视觉空间同时进行语义映射建立和域适应的直推式零样本学习方法容易陷入“相互制衡”问题,从而无法充分发挥语义映射和域适应的最佳性能.针对上述问题,提出了一种基于间接域适应特征生成(feature generation with indirect domain adaptation,FG-IDA)的直推式零样本学习方法.该方法通过串行化语义映射和域适应优化过程,使得直推式零样本学习的这两大核心步骤能够在不同特征空间分别进行最佳优化,从而激发其潜能提升零样本识别精度.在4个标准数据集(CUB,AWA1,AWA2,SUN)上对FG-IDA模型进行了评估,实验结果表明,FG-IDA模型不仅展示出了相对其他直推学习方法的优越性,同时还在AWA1,AWA2和CUB数据集上取得了当前最优结果(the state-of-the-art performance).此外还进行了详尽的消融实验,通过与直接域适应方法进行对比分析,验证了直推式零样本学习中的“相互制衡”问题以及间接域适应思想的先进性.In recent years,zero-shot learning has attracted extensive attention in machine learning and computer vision.The conventional inductive zero-shot learning attempts to establish the mappings between semantic and visual features for transferring the knowledge between classes.However,such approaches suffer from the projection domain shift between the seen and unseen classes.The transductive zero-shot learning is proposed to alleviate this issue by leveraging the unlabeled unseen data for domain adaptation in the training stage.Unfortunately,empirically study finds that these transductive zero-shot learning approaches,which optimize the semantic mapping and domain adaption in visual feature space simultaneously,are easy to trap in“mutual restriction”,and thereby limit the potentials of both these two steps.In order to address the aforementioned issue,a novel transductive zero-shot learning approach named feature generation with indirect domain adaption(FG-IDA)is proposed,that conducts the semantic mapping and domain adaption orderly and optimizes these two steps in different spaces separately for inspiring their performance potentials and further improving the zero-shot recognition accuracy.FG-IDA is evaluated on four benchmarks,namely CUB,AWA1,AWA2,and SUN.The experimental results demonstrate the superiority of the proposed method over other transductive zero-shot learning approaches,and also show that FG-IDA achieves the state-of-the-art performances on CUB,AWA1,and AWA2 datasets.Moreover,the detailed ablation analysis is conducted and the results empirically prove the existence of the“mutual restriction”effect in direct domain adaption-based transductive zero-shot learning approaches and the effectiveness of the indirect domain adaption idea.

关 键 词：图像分类 零样本学习 生成对抗网络 域适应 特征生成 

分 类 号：TP181[自动化与计算机技术—控制理论与控制工程]