面向开放集图像分类的模糊域自适应方法

目前大多的域自适应算法在源域与目标域具有相同类别的场景下,利用标签丰富的源域信息对标签稀少且分布相似的目标域数据进行迁移学习,取得了很多成果。然而,由于现实场景的复杂性和开放性,源域和目标域在类别空间上不尽相同,往往会各自包含一些类别未知且超出现有类别设定的样本。对于这样具有挑战性的开放集场景,传统的域自适应算法将无能为力。为了有效解决上述问题,提出一种面向开放集的模糊域自适应算法。该算法引用了不确定性的模糊化,计算目标域样本的模糊隶属度来学习源域特征到目标域特征空间的线性映射,通过迭代逐步将源域与目标域转化在同一特征空间下。通过对无监督和半监督的图像迁移任务的大量实验,验证了该算法对于开放集场景下图像分类的有效性。

Fuzzy Domain Adaptation Algorithm for Open Set Image Classification

Most of the existing domain adaptation algorithms have been proposed to transfer the knowledge from a welllabeled source domain to scarcely labeled target domain,under the assumption that both source and target domains have different yet similar distributions of training data with the identical class labels,and a lot of achievements are acquired.However,due to the complexity and openness of realistic scenarios,the source domain and the target domain are usually different in category space,and both contain some samples with unknown classes and beyond the existing class settings individually.Traditional domain adaptive algorithms are powerless in this challenging open-set scenario.In order to solve the above problems effectively,this paper proposes a fuzzy domain adaptation algorithm for open sets.This algorithm refers to the fuzziness of uncertainty,learns a linear mapping of source domain features to target domain feature space by calculating the fuzzy membership of target domain samples,and transforms the source domain into the target domain feature space gradually by iteration.Through a large number of experiments on unsupervised and semi-supervised image migration tasks,the effectiveness of this algorithm for open-set image classification is verified.