结合对抗网络与条件均值的多源适应分类方法

Multi-source adaptation classification method combined with adversarial network and conditional mean

生成适应网络利用对抗训练辅助模型进行域适应分类,但仅使用单源域学到的知识有限,且对抗训练不足以减少域差异,造成判别特征难以识别,影响分类精度。针对该问题,提出一种结合对抗网络与条件均值的多源适应分类方法(MSDACG)。对多个源域进行特征提取,提升特征学习的有效部分,对不同源和目标域特征使用特定域的生成对抗网络及条件最大均值差异拉近域间距离,采用差异损失约束由不同源域训练的分类器,实现利用多个源域的监督信息对目标域样本进行分类。实验结果表明,MSDACG模型能学到更优的域不变特征,与目前多源域适应算法比较,其图像生成质量和分类精度有明显提升。

Generate-to-adapt uses adversarial training to assist model in performing domain adaptation classification.However,the knowledge learned by only using a single source domain is limited,and the adversarial training is not enough to reduce domain differences,making it difficult to identify discriminative features,thereby affecting the classification accuracy.For solving this problem,a multi-source adaption classification method combined with adversarial network and conditional mean(MSDACG)was proposed.Features from multiple source domains were extracted,the effective part of feature learning was improved.For the features from different source and target domains,a domain specific generative adversarial network and conditional maximum mean discrepancy(CMMD)were used to reduce the domain distance.For the information of decision boundary between different source and target categories,discrepancy loss was used to constrain the output of classifiers trained in different source domains.Therefore,the target domain samples were classified by using the supervision information of multiple source domains.Experimental results show that the MSDACG model can learn better domain-invariant feature.Compared with the current multi-source domain adaptation algorithm,the image generation quality and classification accuracy are significantly improved.